extracted_excerpts.json

{
  "stats": {
    "total": 286,
    "has_contribution_marker": 1,
    "has_background_marker": 0,
    "has_method_marker": 0,
    "extracted_from_contribution": 1,
    "full_excerpt": 285
  },
  "papers": [
    {
      "url": "/papers/llm/algorithm/architecture/attention/2025/06/01/gated-attention-for-large-language-models-non-linearity-sparsity-and-attention-sink-free.html",
      "title": "Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free",
      "raw_excerpt": "本文系统性地研究了在标准 Softmax 注意力机制中引入门控（gating）机制的影响。",
      "extracted_excerpt": "本文系统性地研究了在标准 Softmax 注意力机制中引入门控（gating）机制的影响。",
      "excerpt_length": 45,
      "extracted_length": 45,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2023/12/01/gated-linear-attention-transformers-with-hardware-efficient-training.html",
      "title": "Gated Linear Attention Transformers with Hardware-Efficient Training",
      "raw_excerpt": "本文致力于解决现有线性注意力模型相较于标准Softmax注意力在性能和实际运行速度上的不足。",
      "extracted_excerpt": "本文致力于解决现有线性注意力模型相较于标准Softmax注意力在性能和实际运行速度上的不足。",
      "excerpt_length": 46,
      "extracted_length": 46,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2022/01/01/transformer-quality-in-linear-time.html",
      "title": "Transformer Quality in Linear Time",
      "raw_excerpt": "核心问题: 尽管Transformer模型【35, Vaswani, A. 等人, Attention is all you need, NIPS 201...",
      "extracted_excerpt": "核心问题: 尽管Transformer模型【35, Vaswani, A. 等人, Attention is all you need, NIPS 201...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2022/10/01/the-devil-in-linear-transformer.html",
      "title": "The Devil in Linear Transformer",
      "raw_excerpt": "本文旨在解决现有基于核（kernel-based）的线性 Transformer 相较于标准 Transformer 性能下降的问题。作者通过深入分析，指...",
      "extracted_excerpt": "本文旨在解决现有基于核（kernel-based）的线性 Transformer 相较于标准 Transformer 性能下降的问题。作者通过深入分析，指...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2025/08/01/artificial-hippocampus-networks-for-efficient-long-context-modeling.html",
      "title": "Artificial Hippocampus Networks for Efficient Long-Context Modeling",
      "raw_excerpt": "核心问题: 长序列建模面临一个根本性的权衡：一方面是类RNN模型中压缩性固定大小内存的高效率，另一方面是基于注意力机制的Transformer中无损增长内...",
      "extracted_excerpt": "核心问题: 长序列建模面临一个根本性的权衡：一方面是类RNN模型中压缩性固定大小内存的高效率，另一方面是基于注意力机制的Transformer中无损增长内...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2021/06/01/linear-transformers-are-secretly-fast-weight-programmers.html",
      "title": "Linear Transformers Are Secretly Fast Weight Programmers",
      "raw_excerpt": "本文的核心在于揭示了线性化自注意力机制与20世纪90年代初的快速权重编程器（Fast Weight Programmers, FWP）在形式上的等价性。基...",
      "extracted_excerpt": "本文的核心在于揭示了线性化自注意力机制与20世纪90年代初的快速权重编程器（Fast Weight Programmers, FWP）在形式上的等价性。基...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2024/04/01/linear-attention-sequence-parallelism.html",
      "title": "Linear Attention Sequence Parallelism",
      "raw_excerpt": "本文针对线性序列建模方法（如线性注意力）提出了一种名为线性注意力序列并行（Linear Attention Sequence Parallelism, L...",
      "extracted_excerpt": "本文针对线性序列建模方法（如线性注意力）提出了一种名为线性注意力序列并行（Linear Attention Sequence Parallelism, L...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2021/02/01/learning-associative-inference-using-fast-weight-memory.html",
      "title": "LEARNING ASSOCIATIVE INFERENCE USING FAST WEIGHT MEMORY",
      "raw_excerpt": "核心问题：\n现代深度神经网络（NNs）尽管在许多人工智能问题上取得了成功，但在需要组合不同经验中提取的特征并进行关联（即组合泛化）的情境中表现不佳。例如，...",
      "extracted_excerpt": "核心问题：\n现代深度神经网络（NNs）尽管在许多人工智能问题上取得了成功，但在需要组合不同经验中提取的特征并进行关联（即组合泛化）的情境中表现不佳。例如，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2025/01/01/minimax-01-scaling-foundation-models-with-lightning-attention.html",
      "title": "MiniMax-01: Scaling Foundation Models with Lightning Attention",
      "raw_excerpt": "本文旨在构建一个性能媲美顶尖商业模型，同时上下文窗口长度大一个数量级的模型。这一目标需要在网络架构、数据和计算之间进行仔细权衡。",
      "extracted_excerpt": "本文旨在构建一个性能媲美顶尖商业模型，同时上下文窗口长度大一个数量级的模型。这一目标需要在网络架构、数据和计算之间进行仔细权衡。",
      "excerpt_length": 64,
      "extracted_length": 64,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2024/05/01/various-lengths-constant-speed-efficient-language-modeling-with-lightning-attention.html",
      "title": "Various Lengths, Constant Speed: Efficient Language Modeling with Lightning Attention",
      "raw_excerpt": "本文旨在解决现有线性注意力机制在大型语言模型中未被广泛采用的两个核心问题：1) 性能不佳：与顶尖的基于Softmax注意力的模型相比，存在明显的性能差距；...",
      "extracted_excerpt": "本文旨在解决现有线性注意力机制在大型语言模型中未被广泛采用的两个核心问题：1) 性能不佳：与顶尖的基于Softmax注意力的模型相比，存在明显的性能差距；...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2023/07/01/scaling-transnormer-to-175-billion-parameters.html",
      "title": "Scaling TransNormer to 175 Billion Parameters",
      "raw_excerpt": "本文旨在解决传统Transformer模型中softmax注意力机制带来的二次方时间复杂度问题，该问题限制了模型在训练和推理阶段的可扩展性和效率。尽管已有...",
      "extracted_excerpt": "本文旨在解决传统Transformer模型中softmax注意力机制带来的二次方时间复杂度问题，该问题限制了模型在训练和推理阶段的可扩展性和效率。尽管已有...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2024/01/01/lightning-attention-2-a-free-lunch-for-handling-unlimited-sequence-lengths-in-large-language-models.html",
      "title": "Lightning Attention-2: A Free Lunch for Handling Unlimited Sequence Lengths in Large Language Models",
      "raw_excerpt": "核心问题\nTransformer架构的计算复杂度随着输入序列的长度呈二次方增长，这使得处理极长序列变得具有挑战性。尽管线性注意力理论上通过核技巧可以将计算...",
      "extracted_excerpt": "核心问题\nTransformer架构的计算复杂度随着输入序列的长度呈二次方增长，这使得处理极长序列变得具有挑战性。尽管线性注意力理论上通过核技巧可以将计算...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/linear/2020/06/01/transformers-are-rnns-fast-autoregressive-transformers-with-linear-attention.html",
      "title": "Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention",
      "raw_excerpt": "核心问题：\n标准的Transformer模型虽然在多种任务中表现出色，但其核心组件自注意力（self-attention）的计算和内存复杂性与输入序列长度...",
      "extracted_excerpt": "核心问题：\n标准的Transformer模型虽然在多种任务中表现出色，但其核心组件自注意力（self-attention）的计算和内存复杂性与输入序列长度...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2021/10/01/combining-recurrent-convolutional-and-continuous-time-models-with-linear-state-space-layers.html",
      "title": "Combining Recurrent, Convolutional, and Continuous-time Models with Linear State-Space Layers",
      "raw_excerpt": "本文旨在解决机器学习中高效建模长序列数据（超过几千个时间步）的挑战。传统序列模型范式，如循环神经网络（RNNs）、卷积神经网络（CNNs）和神经微分方程（...",
      "extracted_excerpt": "本文旨在解决机器学习中高效建模长序列数据（超过几千个时间步）的挑战。传统序列模型范式，如循环神经网络（RNNs）、卷积神经网络（CNNs）和神经微分方程（...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2019/01/01/transformer-xl-attentive-language-models-beyond-a-fixed-length-context.html",
      "title": "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
      "raw_excerpt": "核心问题：尽管Transformer在学习长期依赖方面具有潜力，但在语言建模中，它们受到固定长度上下文的限制。这种方法存在两个关键问题：1) 模型无法捕获...",
      "extracted_excerpt": "核心问题：尽管Transformer在学习长期依赖方面具有潜力，但在语言建模中，它们受到固定长度上下文的限制。这种方法存在两个关键问题：1) 模型无法捕获...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2024/12/01/gated-delta-networks-improving-mamba2-with-delta-rule.html",
      "title": "GATED DELTA NETWORKS: IMPROVING MAMBA2 WITH DELTA RULE",
      "raw_excerpt": "本文旨在解决线性Transformer在检索和长上下文任务中性能受限的问题。尽管线性Transformer作为标准Transformer的高效替代品备受关...",
      "extracted_excerpt": "本文旨在解决线性Transformer在检索和长上下文任务中性能受限的问题。尽管线性Transformer作为标准Transformer的高效替代品备受关...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2024/03/01/jamba-a-hybrid-transformer-mamba-language-model.html",
      "title": "Jamba: A Hybrid Transformer-Mamba Language Model",
      "raw_excerpt": "本文介绍了Jamba，一种新型的、公开可用的大型语言模型。Jamba基于一种创新的混合架构，该架构融合了Transformer层、Mamba层（一种最新的...",
      "extracted_excerpt": "本文介绍了Jamba，一种新型的、公开可用的大型语言模型。Jamba基于一种创新的混合架构，该架构融合了Transformer层、Mamba层（一种最新的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2023/12/01/mamba-linear-time-sequence-modeling-with-selective-state-spaces.html",
      "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces",
      "raw_excerpt": "本文旨在解决现有主流基础模型（Foundation Models）骨干架构 Transformer 在处理长序列时计算效率低下的问题。Transforme...",
      "extracted_excerpt": "本文旨在解决现有主流基础模型（Foundation Models）骨干架构 Transformer 在处理长序列时计算效率低下的问题。Transforme...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2024/05/01/transformers-are-ssms-generalized-models-and-efficient-algorithms-through-structured-state-space-duality.html",
      "title": "Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality",
      "raw_excerpt": "本文的核心目标是建立结构化状态空间模型（SSMs）与多种注意力变体之间的丰富理论联系，从而将为Transformer开发的算法和系统优化迁移到SSM上，构...",
      "extracted_excerpt": "本文的核心目标是建立结构化状态空间模型（SSMs）与多种注意力变体之间的丰富理论联系，从而将为Transformer开发的算法和系统优化迁移到SSM上，构...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2023/04/01/hungry-hungry-hippos-towards-language-modeling-with-state-space-models.html",
      "title": "Hungry Hungry Hippos: Towards Language Modeling with State Space Models",
      "raw_excerpt": "本文旨在解决状态空间模型（SSM）在语言建模领域相较于Transformer存在的两个核心问题：模型表达能力不足和硬件利用率低下导致的训练速度慢。",
      "extracted_excerpt": "本文旨在解决状态空间模型（SSM）在语言建模领域相较于Transformer存在的两个核心问题：模型表达能力不足和硬件利用率低下导致的训练速度慢。",
      "excerpt_length": 73,
      "extracted_length": 73,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2021/10/01/efficiently-modeling-long-sequences-with-structured-state-spaces.html",
      "title": "Efficiently Modeling Long Sequences with Structured State Spaces",
      "raw_excerpt": "本文旨在解决序列建模中的一个核心问题：高效处理包含长距离依赖（LRDs）的数据。现有的主流模型，如循环神经网络（RNNs）、卷积神经网络（CNNs）和Tr...",
      "extracted_excerpt": "本文旨在解决序列建模中的一个核心问题：高效处理包含长距离依赖（LRDs）的数据。现有的主流模型，如循环神经网络（RNNs）、卷积神经网络（CNNs）和Tr...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2023/03/01/simplified-state-space-layers-for-sequence-modeling.html",
      "title": "SIMPLIFIED STATE SPACE LAYERS FOR SEQUENCE MODELING",
      "raw_excerpt": "本文旨在解决机器学习中高效建模长序列的挑战性问题，即序列中相隔数千个时间步的观测值可能共同编码了解决任务的关键信息。尽管已有利普希茨循环神经网络（RNN）...",
      "extracted_excerpt": "本文旨在解决机器学习中高效建模长序列的挑战性问题，即序列中相隔数千个时间步的观测值可能共同编码了解决任务的关键信息。尽管已有利普希茨循环神经网络（RNN）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/ssm/2024/02/01/moe-mamba-efficient-selective-state-space-models-with-mixture-of-experts.html",
      "title": "MoE-Mamba: Efficient Selective State Space Models with Mixture of Experts",
      "raw_excerpt": "本文旨在解决现有大型语言模型（LLM）架构的局限性，并探索进一步扩展语言模型的可能性。当前LLM主要依赖于Transformer架构【65, Vaswan...",
      "extracted_excerpt": "本文旨在解决现有大型语言模型（LLM）架构的局限性，并探索进一步扩展语言模型的可能性。当前LLM主要依赖于Transformer架构【65, Vaswan...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/sparsity/2020/04/01/longformer-the-long-document-transformer.html",
      "title": "Longformer: The Long-Document Transformer",
      "raw_excerpt": "本文旨在解决基于Transformer的模型因其自注意力操作（其计算量与序列长度成二次方关系）而无法处理长序列的局限性。",
      "extracted_excerpt": "本文旨在解决基于Transformer的模型因其自注意力操作（其计算量与序列长度成二次方关系）而无法处理长序列的局限性。",
      "excerpt_length": 60,
      "extracted_length": 60,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/sparsity/2025/02/01/moba-mixture-of-block-attention-for-long-context-llms.html",
      "title": "MOBA: MIXTURE OF BLOCK ATTENTION FOR LONG-CONTEXT LLMS",
      "raw_excerpt": "本文旨在解决大语言模型（LLM）在处理长序列时面临的核心挑战，即传统自注意力机制带来的二次方计算复杂度增长问题。现有方法通常引入强结构偏见（如窗口注意力）...",
      "extracted_excerpt": "本文旨在解决大语言模型（LLM）在处理长序列时面临的核心挑战，即传统自注意力机制带来的二次方计算复杂度增长问题。现有方法通常引入强结构偏见（如窗口注意力）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/architecture/attention/sparsity/2025/02/01/native-sparse-attention-hardware-aligned-and-natively-trainable-sparse-attention.html",
      "title": "Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention",
      "raw_excerpt": "本文提出了一种名为NSA（Natively trainable Sparse Attention，原生可训练稀疏注意力）的机制，旨在解决长上下文语言模型中...",
      "extracted_excerpt": "本文提出了一种名为NSA（Natively trainable Sparse Attention，原生可训练稀疏注意力）的机制，旨在解决长上下文语言模型中...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2025/04/01/mem0-building-production-ready-ai-agents-with-scalable-long-term-memory.html",
      "title": "Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory",
      "raw_excerpt": "本文旨在解决大型语言模型（LLMs）因固定上下文窗口而在跨会话的长时间对话中难以保持连贯性的根本问题。为应对这一挑战，研究者引入了名为 Mem0 的新型记...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLMs）因固定上下文窗口而在跨会话的长时间对话中难以保持连贯性的根本问题。为应对这一挑战，研究者引入了名为 Mem0 的新型记...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2025/03/01/search-r1-training-llms-to-reason-and-leverage-search-engines-with-reinforcement-learning.html",
      "title": "Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning",
      "raw_excerpt": "本文旨在解决大型语言模型（LLMs）在处理复杂推理和获取最新外部信息时面临的挑战。现有的方法，如检索增强生成（RAG）或将搜索引擎视为工具的提示方法，通常...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLMs）在处理复杂推理和获取最新外部信息时面临的挑战。现有的方法，如检索增强生成（RAG）或将搜索引擎视为工具的提示方法，通常...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2022/01/01/chain-of-thought-prompting-elicits-reasoning-in-large-language-models.html",
      "title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models",
      "raw_excerpt": "本文探讨了一种名为“思维链（Chain-of-Thought, CoT）”提示的方法，该方法通过生成一系列中间推理步骤，显著提高了大型语言模型执行复杂推理...",
      "extracted_excerpt": "本文探讨了一种名为“思维链（Chain-of-Thought, CoT）”提示的方法，该方法通过生成一系列中间推理步骤，显著提高了大型语言模型执行复杂推理...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2024/09/01/large-language-model-based-agents-for-software-engineering-a-survey.html",
      "title": "Large Language Model-Based Agents for Software Engineering: A Survey",
      "raw_excerpt": "本文首次对106篇将基于大型语言模型（LLM）的智能体应用于软件工程（SE）的论文进行了全面综述。研究从软件工程和智能体两个视角，分析了现有基于LLM的智...",
      "extracted_excerpt": "本文首次对106篇将基于大型语言模型（LLM）的智能体应用于软件工程（SE）的论文进行了全面综述。研究从软件工程和智能体两个视角，分析了现有基于LLM的智...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/10/01/fireact-toward-language-agent-fine-tuning.html",
      "title": "FIREACT: TOWARD LANGUAGE AGENT FINE-TUNING",
      "raw_excerpt": "本文探讨了通过微调语言模型（LMs）来构建语言代理这一被忽视的研究方向。",
      "extracted_excerpt": "本文探讨了通过微调语言模型（LMs）来构建语言代理这一被忽视的研究方向。",
      "excerpt_length": 36,
      "extracted_length": 36,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/12/01/retrieval-augmented-generation-for-large-language-models-a-survey.html",
      "title": "Retrieval-Augmented Generation for Large Language Models: A Survey",
      "raw_excerpt": "本文对检索增强生成（RAG）领域进行了系统性的回顾和梳理。其主要贡献如下：",
      "extracted_excerpt": "",
      "excerpt_length": 37,
      "extracted_length": 0,
      "is_extracted": true
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/11/01/reac-t-synergizing-reasoning-and-acting-in-language-models.html",
      "title": "REAC T: SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS",
      "raw_excerpt": "本文探讨了如何利用大型语言模型（LLMs）以交错的方式生成推理轨迹和任务特定动作，从而实现两者之间更大的协同作用。核心问题在于，现有工作中，LLMs 的推...",
      "extracted_excerpt": "本文探讨了如何利用大型语言模型（LLMs）以交错的方式生成推理轨迹和任务特定动作，从而实现两者之间更大的协同作用。核心问题在于，现有工作中，LLMs 的推...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2025/09/01/the-landscape-of-agentic-reinforcement-learning-for-llms-a-survey.html",
      "title": "The Landscape of Agentic Reinforcement Learning for LLMs: A Survey",
      "raw_excerpt": "本文系统性地综述了“智能体强化学习”（Agentic RL）这一新兴范式，该范式将大型语言模型（LLM）从被动的序列生成器转变为嵌入在复杂动态世界中的自主...",
      "extracted_excerpt": "本文系统性地综述了“智能体强化学习”（Agentic RL）这一新兴范式，该范式将大型语言模型（LLM）从被动的序列生成器转变为嵌入在复杂动态世界中的自主...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/02/01/toolformer-language-models-can-teach-themselves-to-use-tools.html",
      "title": "Toolformer: Language Models Can Teach Themselves to Use Tools",
      "raw_excerpt": "核心问题: 大型语言模型（LLMs）虽然在许多自然语言处理任务上表现出色，但仍存在一些固有限制，即使通过进一步扩大模型规模也难以完全解决。这些限制包括：\n...",
      "extracted_excerpt": "核心问题: 大型语言模型（LLMs）虽然在许多自然语言处理任务上表现出色，但仍存在一些固有限制，即使通过进一步扩大模型规模也难以完全解决。这些限制包括：\n...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2025/03/01/r1-searcher-incentivizing-the-search-capability-in-llms-via-reinforcement-learning.html",
      "title": "R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning",
      "raw_excerpt": "本文旨在解决现有大型推理模型（LRM）在处理对时间敏感或知识密集型问题时，因依赖内部知识而导致的不准确和幻觉问题。尽管检索增强生成（RAG）技术被广泛研究...",
      "extracted_excerpt": "本文旨在解决现有大型推理模型（LRM）在处理对时间敏感或知识密集型问题时，因依赖内部知识而导致的不准确和幻觉问题。尽管检索增强生成（RAG）技术被广泛研究...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2025/05/01/a-survey-on-test-time-scaling-in-large-language-models-what-how-where-and-how-well.html",
      "title": "A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well",
      "raw_excerpt": "核心问题：随着预训练阶段通过扩展计算（数据和参数）带来的性能提升逐渐放缓，研究重心已转向如何在测试时充分激发大型语言模型（LLMs）中编码的智能，以最大化...",
      "extracted_excerpt": "核心问题：随着预训练阶段通过扩展计算（数据和参数）带来的性能提升逐渐放缓，研究重心已转向如何在测试时充分激发大型语言模型（LLMs）中编码的智能，以最大化...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/03/01/reac-t-synergizing-reasoning-and-acting-in-language-models.html",
      "title": "REAC T: SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS",
      "raw_excerpt": "本文探讨了如何利用大型语言模型（LLMs）以交错的方式生成推理轨迹和任务特定动作，从而实现两者之间更强的协同作用。核心问题在于，先前研究主要将LLMs的推...",
      "extracted_excerpt": "本文探讨了如何利用大型语言模型（LLMs）以交错的方式生成推理轨迹和任务特定动作，从而实现两者之间更强的协同作用。核心问题在于，先前研究主要将LLMs的推...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/12/01/tree-of-thoughts-deliberate-problem-solving-with-large-language-models.html",
      "title": "Tree of Thoughts: Deliberate Problem Solving with Large Language Models",
      "raw_excerpt": "本文旨在解决大型语言模型（LMs）在推理时局限于逐个词元（token-level）、从左到右决策过程的问题，这一机制在需要探索、策略性前瞻或初始决策至关重...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LMs）在推理时局限于逐个词元（token-level）、从左到右决策过程的问题，这一机制在需要探索、策略性前瞻或初始决策至关重...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2022/07/01/webshop-towards-scalable-real-world-web-interaction-with-grounded-language-agents.html",
      "title": "WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents",
      "raw_excerpt": "核心问题与研究目标: 近期的自然语言处理（NLP）和强化学习（RL）在能够利用语言上下文进行序列决策的智能体方面取得了进展。然而，一方面，这些交互式任务在...",
      "extracted_excerpt": "核心问题与研究目标: 近期的自然语言处理（NLP）和强化学习（RL）在能够利用语言上下文进行序列决策的智能体方面取得了进展。然而，一方面，这些交互式任务在...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/agent/2023/09/01/tree-of-thoughts-deliberate-problem-solving-with-large-language-models.html",
      "title": "Tree of Thoughts: Deliberate Problem Solving with Large Language Models",
      "raw_excerpt": "本文旨在解决大型语言模型（LMs）在推理时局限于逐个词元（token-level）、从左到右决策过程的问题，这一机制在需要探索、策略性前瞻或初始决策至关重...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LMs）在推理时局限于逐个词元（token-level）、从左到右决策过程的问题，这一机制在需要探索、策略性前瞻或初始决策至关重...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2023/05/01/direct-preference-optimization-your-language-model-is-secretly-a-reward-model.html",
      "title": "Direct Preference Optimization: Your Language Model is Secretly a Reward Model",
      "raw_excerpt": "图1：DPO在避免强化学习的同时优化人类偏好。现有的使用人类反馈微调语言模型的方法首先将奖励模型拟合到提示和人类对响应对的偏好数据集上，然后使用RL找到一...",
      "extracted_excerpt": "图1：DPO在避免强化学习的同时优化人类偏好。现有的使用人类反馈微调语言模型的方法首先将奖励模型拟合到提示和人类对响应对的偏好数据集上，然后使用RL找到一...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2016/01/01/mastering-the-game-of-go-with-deep-neural-networks-and-tree-search.html",
      "title": "Mastering the game of Go with deep neural networks and tree search",
      "raw_excerpt": "本文针对围棋游戏的巨大搜索空间和棋盘位置评估难度，提出了一种新的计算机围棋方法，使用“价值网络”评估棋盘位置和“策略网络”选择走法。这些深度神经网络通过从...",
      "extracted_excerpt": "本文针对围棋游戏的巨大搜索空间和棋盘位置评估难度，提出了一种新的计算机围棋方法，使用“价值网络”评估棋盘位置和“策略网络”选择走法。这些深度神经网络通过从...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2025/05/01/the-entropy-mechanism-of-reinforcement-learning-for-reasoning-language-models.html",
      "title": "The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models",
      "raw_excerpt": "本文旨在克服在扩展强化学习（RL）以用于大型语言模型（LLMs）推理时的一个主要障碍，即策略熵的崩溃。",
      "extracted_excerpt": "本文旨在克服在扩展强化学习（RL）以用于大型语言模型（LLMs）推理时的一个主要障碍，即策略熵的崩溃。",
      "excerpt_length": 51,
      "extracted_length": 51,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2024/08/01/scaling-llm-test-time-compute-optimally-can-be-more-effective-than-scaling-model-parameters.html",
      "title": "Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters",
      "raw_excerpt": "本文旨在探讨如何通过增加测试时（test-time）的计算量来提升大型语言模型（LLM）的性能，特别是针对具有挑战性的任务。研究的核心问题是：在给定固定的...",
      "extracted_excerpt": "本文旨在探讨如何通过增加测试时（test-time）的计算量来提升大型语言模型（LLM）的性能，特别是针对具有挑战性的任务。研究的核心问题是：在给定固定的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2022/03/01/training-language-models-to-follow-instructions-with-human-feedback.html",
      "title": "Training language models to follow instructions with human feedback",
      "raw_excerpt": "本文旨在解决大型语言模型（LMs）本质上并不擅长遵循用户意图的问题，即模型与用户不“对齐”（aligned）。大型语言模型可能会生成不真实、有毒或对用户无...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LMs）本质上并不擅长遵循用户意图的问题，即模型与用户不“对齐”（aligned）。大型语言模型可能会生成不真实、有毒或对用户无...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2024/01/01/self-play-fine-tuning-converts-weak-language-models-to-strong-language-models.html",
      "title": "Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models",
      "raw_excerpt": "本文研究了如何在不获取额外人工标注数据的情况下，将一个弱大型语言模型（LLM）提升为一个强LLM。",
      "extracted_excerpt": "本文研究了如何在不获取额外人工标注数据的情况下，将一个弱大型语言模型（LLM）提升为一个强LLM。",
      "excerpt_length": 49,
      "extracted_length": 49,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2024/07/01/helpsteer2-open-source-dataset-for-training-top-performing-reward-models.html",
      "title": "HelpSteer2: Open-source dataset for training top-performing reward models",
      "raw_excerpt": "本文介绍并发布了HelpSteer2，一个高质量、遵循宽松许可（CC-BY-4.0）的偏好数据集，旨在解决当前大型语言模型（LLM）对齐领域中高质量、开放...",
      "extracted_excerpt": "本文介绍并发布了HelpSteer2，一个高质量、遵循宽松许可（CC-BY-4.0）的偏好数据集，旨在解决当前大型语言模型（LLM）对齐领域中高质量、开放...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2017/12/01/rllib-abstractions-for-distributed-reinforcement-learning.html",
      "title": "RLlib: Abstractions for Distributed Reinforcement Learning",
      "raw_excerpt": "本文旨在解决强化学习（RL）领域在系统和抽象设计方面进展缓慢的问题，与深度学习（DL）领域形成鲜明对比。尽管RL社区受益于DL的系统进步，但RL算法固有的...",
      "extracted_excerpt": "本文旨在解决强化学习（RL）领域在系统和抽象设计方面进展缓慢的问题，与深度学习（DL）领域形成鲜明对比。尽管RL社区受益于DL的系统进步，但RL算法固有的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2025/08/01/on-policy-rl-meets-off-policy-experts-harmonizing-supervised-fine-tuning-and-reinforcement-learning-via-dynamic-weighting.html",
      "title": "ON-POLICY RL MEETS OFF-POLICY EXPERTS: HARMONIZING SUPERVISED FINE-TUNING AND REINFORCEMENT LEARNING VIA DYNAMIC WEIGHTING",
      "raw_excerpt": "本文旨在解决整合监督微调（SFT）和强化学习（RL）时遇到的挑战，特别是现有方法可能破坏模型已建立的模式并导致对专家数据的过拟合。",
      "extracted_excerpt": "本文旨在解决整合监督微调（SFT）和强化学习（RL）时遇到的挑战，特别是现有方法可能破坏模型已建立的模式并导致对专家数据的过拟合。",
      "excerpt_length": 65,
      "extracted_length": 65,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2023/10/01/steerlm-attribute-conditioned-sft-as-an-user-steerable-alternative-to-rlhf.html",
      "title": "SteerLM: Attribute Conditioned SFT as an (User-Steerable) Alternative to RLHF",
      "raw_excerpt": "本文旨在解决现有大语言模型（LLM）对齐方法（特别是基于人类反馈的强化学习，RLHF）的局限性。",
      "extracted_excerpt": "本文旨在解决现有大语言模型（LLM）对齐方法（特别是基于人类反馈的强化学习，RLHF）的局限性。",
      "excerpt_length": 48,
      "extracted_length": 48,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2025/05/01/dapo-an-open-source-llm-reinforcement-learning-system-at-scale.html",
      "title": "DAPO: An Open-Source LLM Reinforcement Learning System at Scale",
      "raw_excerpt": "论文的核心问题是大规模强化学习（RL）在大型语言模型（LLM）中的实际算法和关键技巧仍被隐藏，导致社区难以重现现有推理模型的训练结果，如OpenAI o1...",
      "extracted_excerpt": "论文的核心问题是大规模强化学习（RL）在大型语言模型（LLM）中的实际算法和关键技巧仍被隐藏，导致社区难以重现现有推理模型的训练结果，如OpenAI o1...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2022/04/01/training-a-helpful-and-harmless-assistant-with-reinforcement-learning-from-human-feedback.html",
      "title": "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback",
      "raw_excerpt": "本文旨在通过偏好建模（Preference Modeling, PM）和基于人类反馈的强化学习（Reinforcement Learning from H...",
      "extracted_excerpt": "本文旨在通过偏好建模（Preference Modeling, PM）和基于人类反馈的强化学习（Reinforcement Learning from H...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2013/12/01/playing-atari-with-deep-reinforcement-learning.html",
      "title": "Playing Atari with Deep Reinforcement Learning",
      "raw_excerpt": "本文提出首个能够直接从高维感官输入（如原始像素）中，通过强化学习成功学习控制策略的深度学习模型。",
      "extracted_excerpt": "本文提出首个能够直接从高维感官输入（如原始像素）中，通过强化学习成功学习控制策略的深度学习模型。",
      "excerpt_length": 48,
      "extracted_length": 48,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2015/07/01/massively-parallel-methods-for-deep-reinforcement-learning.html",
      "title": "Massively Parallel Methods for Deep Reinforcement Learning",
      "raw_excerpt": "本文介绍了一种首次为深度强化学习设计的大规模分布式架构。由于先前最先进的深度强化学习算法，如深度Q网络（DQN），仅在单机架构上应用，导致训练时间过长（例...",
      "extracted_excerpt": "本文介绍了一种首次为深度强化学习设计的大规模分布式架构。由于先前最先进的深度强化学习算法，如深度Q网络（DQN），仅在单机架构上应用，导致训练时间过长（例...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2025/04/01/exploring-data-scaling-trends-and-effects-in-reinforcement-learning-from-human-feedback.html",
      "title": "Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback",
      "raw_excerpt": "本文探讨了人类反馈强化学习（RLHF）中数据缩放的趋势和影响，重点关注当前阻碍RLHF性能缩放的数据驱动瓶颈，特别是奖励黑客攻击和响应多样性下降的问题。研...",
      "extracted_excerpt": "本文探讨了人类反馈强化学习（RLHF）中数据缩放的趋势和影响，重点关注当前阻碍RLHF性能缩放的数据驱动瓶颈，特别是奖励黑客攻击和响应多样性下降的问题。研...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2017/08/01/proximal-policy-optimization-algorithms.html",
      "title": "Proximal Policy Optimization Algorithms",
      "raw_excerpt": "核心问题：现有的主流强化学习方法存在各自的缺陷。深度Q学习（Deep Q-learning）在许多简单问题上会失败；“香草”策略梯度方法（vanilla ...",
      "extracted_excerpt": "核心问题：现有的主流强化学习方法存在各自的缺陷。深度Q学习（Deep Q-learning）在许多简单问题上会失败；“香草”策略梯度方法（vanilla ...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2024/02/01/deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.html",
      "title": "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models",
      "raw_excerpt": "本文介绍了 DeepSeekMath，一个在数学推理方面取得显著性能的领域特定语言模型。其核心贡献可分为可扩展的数学预训练和对强化学习的探索与分析两个方面。",
      "extracted_excerpt": "本文介绍了 DeepSeekMath，一个在数学推理方面取得显著性能的领域特定语言模型。其核心贡献可分为可扩展的数学预训练和对强化学习的探索与分析两个方面。",
      "excerpt_length": 78,
      "extracted_length": 78,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2020/09/01/learning-to-summarize-from-human-feedback.html",
      "title": "Learning to summarize from human feedback",
      "raw_excerpt": "本文的核心问题是，当前用于微调大型语言模型的监督学习目标（即最大化人类编写文本的对数概率）与我们真正关心的目标（即生成由人类判断的高质量输出）之间存在偏差...",
      "extracted_excerpt": "本文的核心问题是，当前用于微调大型语言模型的监督学习目标（即最大化人类编写文本的对数概率）与我们真正关心的目标（即生成由人类判断的高质量输出）之间存在偏差...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2024/05/01/nemo-aligner-scalable-toolkit-for-efficient-model-alignment.html",
      "title": "NeMo-Aligner: Scalable Toolkit for Efficient Model Alignment",
      "raw_excerpt": "本文介绍了NeMo-Aligner，一个旨在高效地将大型语言模型（LLMs）与人类价值观和偏好对齐的工具包。对齐是使LLMs变得有用和安全的关键步骤，但为...",
      "extracted_excerpt": "本文介绍了NeMo-Aligner，一个旨在高效地将大型语言模型（LLMs）与人类价值观和偏好对齐的工具包。对齐是使LLMs变得有用和安全的关键步骤，但为...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2025/05/01/prorl-prolonged-reinforcement-learning-expands-reasoning-boundaries-in-large-language-models.html",
      "title": "ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models",
      "raw_excerpt": "本文探讨了强化学习（RL）在提升大型语言模型（LLM）推理能力方面的核心问题：RL究竟是解锁了模型新的推理能力，还是仅仅优化了基础模型中已存在的解决方案的...",
      "extracted_excerpt": "本文探讨了强化学习（RL）在提升大型语言模型（LLM）推理能力方面的核心问题：RL究竟是解锁了模型新的推理能力，还是仅仅优化了基础模型中已存在的解决方案的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/rl/2025/07/01/group-sequence-policy-optimization.html",
      "title": "Group Sequence Policy Optimization",
      "raw_excerpt": "本文针对现有强化学习（RL）算法在训练大型语言模型时（特别是GRPO算法）存在的严重稳定性问题，提出了组序列策略优化（Group Sequence Pol...",
      "extracted_excerpt": "本文针对现有强化学习（RL）算法在训练大型语言模型时（特别是GRPO算法）存在的严重稳定性问题，提出了组序列策略优化（Group Sequence Pol...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2025/03/01/samplemix-a-sample-wise-pre-training-data-mixing-strategey-by-coordinating-data-quality-and-diversity.html",
      "title": "SampleMix: A Sample-wise Pre-training Data Mixing Strategey by Coordinating Data Quality and Diversity",
      "raw_excerpt": "本文研究了大型语言模型（LLM）预训练数据的混合策略问题。现有的数据混合方法通常遵循一种“领域为单位”（domain-wise）的自顶向下方法，即先确定各...",
      "extracted_excerpt": "本文研究了大型语言模型（LLM）预训练数据的混合策略问题。现有的数据混合方法通常遵循一种“领域为单位”（domain-wise）的自顶向下方法，即先确定各...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2021/04/01/roformer-enhanced-transformer-with-rotary-position-embedding.html",
      "title": "ROFORMER: ENHANCED TRANSFORMER WITH ROTARY POSITION EMBEDDING",
      "raw_excerpt": "本文研究了将位置信息集成到基于Transformer的语言模型中的多种方法，并提出了一种名为旋转位置嵌入（Rotary Position Embeddin...",
      "extracted_excerpt": "本文研究了将位置信息集成到基于Transformer的语言模型中的多种方法，并提出了一种名为旋转位置嵌入（Rotary Position Embeddin...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2024/12/01/unveiling-the-secret-recipe-a-guide-for-supervised-fine-tuning-small-llms.html",
      "title": "UNVEILING THE SECRET RECIPE: A GUIDE FOR SUPERVISED FINE-TUNING SMALL LLMS",
      "raw_excerpt": "本文旨在弥合大型工业研究实验室与资源有限的个人开发者及小型组织之间在微调大型语言模型（LLM）方面的差距。研究的核心问题是：如何有效地在涵盖多样化知识和技...",
      "extracted_excerpt": "本文旨在弥合大型工业研究实验室与资源有限的个人开发者及小型组织之间在微调大型语言模型（LLM）方面的差距。研究的核心问题是：如何有效地在涵盖多样化知识和技...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2022/03/01/tensor-programs-v-tuning-large-neural-networks-via-zero-shot-hyperparameter-transfer.html",
      "title": "Tensor Programs V: Tuning Large Neural Networks via Zero-Shot\n      Hyperparameter Transfer",
      "raw_excerpt": "核心问题：深度学习中的超参数（HP）调优是一个昂贵的过程，对于具有数十亿参数的神经网络（NNs）而言尤其如此，因为最先进的网络训练成本极高，导致调优变得不...",
      "extracted_excerpt": "核心问题：深度学习中的超参数（HP）调优是一个昂贵的过程，对于具有数十亿参数的神经网络（NNs）而言尤其如此，因为最先进的网络训练成本极高，导致调优变得不...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2022/03/01/deepnet-scaling-transformers-to-1000-layers.html",
      "title": "DeepNet: Scaling Transformers to 1,000 Layers",
      "raw_excerpt": "核心问题：尽管Transformer模型参数量已达万亿级别，但其深度受限于训练不稳定性，通常不超过数百层。现有的稳定化方法，如Pre-LN，虽然提高了稳定...",
      "extracted_excerpt": "核心问题：尽管Transformer模型参数量已达万亿级别，但其深度受限于训练不稳定性，通常不超过数百层。现有的稳定化方法，如Pre-LN，虽然提高了稳定...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2024/08/01/inference-scaling-laws-an-empirical-analysis-of-compute-optimal-inference-for-llm-problem-solving.html",
      "title": "INFERENCE SCALING LAWS: AN EMPIRICAL ANALYSIS OF COMPUTE-OPTIMAL INFERENCE FOR LLM PROBLEM-SOLVING",
      "raw_excerpt": "本文研究了大型语言模型（LLM）的推理缩放定律（或称测试时缩放定律）和计算最优推理，重点探讨了模型大小与不同推理策略下生成额外token之间的权衡。",
      "extracted_excerpt": "本文研究了大型语言模型（LLM）的推理缩放定律（或称测试时缩放定律）和计算最优推理，重点探讨了模型大小与不同推理策略下生成额外token之间的权衡。",
      "excerpt_length": 74,
      "extracted_length": 74,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2013/12/01/what-makes-good-data-for-alignment-a-comprehensive-study-of-automatic-data-selection-in-instruction-tuning.html",
      "title": "WHAT MAKES GOOD DATA FOR ALIGNMENT? A COMPREHENSIVE STUDY OF AUTOMATIC DATA SELECTION IN INSTRUCTION TUNING",
      "raw_excerpt": "本文旨在系统性地探究指令微调中“好数据”的特征，并基于此提出一种自动、高效的数据选择方法，以提升大型语言模型（LLM）对齐的效率。",
      "extracted_excerpt": "本文旨在系统性地探究指令微调中“好数据”的特征，并基于此提出一种自动、高效的数据选择方法，以提升大型语言模型（LLM）对齐的效率。",
      "excerpt_length": 65,
      "extracted_length": 65,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/pretrain_sft/2024/05/01/stacking-your-transformers-a-closer-look-at-model-growth-for-efficient-llm-pre-training.html",
      "title": "Stacking Your Transformers: A Closer Look at Model Growth for Efficient LLM Pre-Training",
      "raw_excerpt": "本文旨在解决大语言模型（LLM）预训练成本高昂的问题，通过一种名为“模型增长”的方法来加速训练过程。模型增长的核心思想是利用已训练好的小模型来初始化并加速...",
      "extracted_excerpt": "本文旨在解决大语言模型（LLM）预训练成本高昂的问题，通过一种名为“模型增长”的方法来加速训练过程。模型增长的核心思想是利用已训练好的小模型来初始化并加速...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/01/01/kimi-k15-scaling-reinforcement-learning-with-llms.html",
      "title": "KIMI K1.5: SCALING REINFORCEMENT LEARNING WITH LLMS",
      "raw_excerpt": "本文介绍了Kimi k1.5的训练方法，这是一个利用强化学习（RL）进行训练的多模态大语言模型（LLM），旨在探索超越现有静态数据集限制的持续扩展新路径。...",
      "extracted_excerpt": "本文介绍了Kimi k1.5的训练方法，这是一个利用强化学习（RL）进行训练的多模态大语言模型（LLM），旨在探索超越现有静态数据集限制的持续扩展新路径。...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/09/01/mimo-audio-audio-language-models-are-few-shot-learners.html",
      "title": "MiMo-Audio: Audio Language Models are Few-Shot Learners",
      "raw_excerpt": "本文的核心论点是，通过大规模的下一词元预测预训练，可以像文本领域的GPT-3一样，在音频领域实现任务的泛化能力。现有的音频语言模型通常依赖于针对特定音频任...",
      "extracted_excerpt": "本文的核心论点是，通过大规模的下一词元预测预训练，可以像文本领域的GPT-3一样，在音频领域实现任务的泛化能力。现有的音频语言模型通常依赖于针对特定音频任...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/06/01/dotsllm1-technical-report.html",
      "title": "dots.llm1 Technical Report",
      "raw_excerpt": "本文介绍了dots.llm1，这是一款大规模、高性价比的专家混合（Mixture of Experts, MoE）模型。该模型总参数量为1420亿，但每个...",
      "extracted_excerpt": "本文介绍了dots.llm1，这是一款大规模、高性价比的专家混合（Mixture of Experts, MoE）模型。该模型总参数量为1420亿，但每个...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/09/01/deepseek-v32-exp-boosting-long-context-efficiency-with-deepseek-sparse-attention.html",
      "title": "DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention",
      "raw_excerpt": "本文介绍了 DeepSeek-V3.2-Exp，这是一个实验性的稀疏注意力模型。该模型通过在 DeepSeek-V3.1-Terminus 的基础上进行持...",
      "extracted_excerpt": "本文介绍了 DeepSeek-V3.2-Exp，这是一个实验性的稀疏注意力模型。该模型通过在 DeepSeek-V3.1-Terminus 的基础上进行持...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/09/01/longcat-flash-thinking-technical-report.html",
      "title": "LongCat-Flash-Thinking Technical Report",
      "raw_excerpt": "本文介绍了 LongCat-Flash-Thinking，一个高效的、拥有 5600 亿参数的开源混合专家（MoE）推理模型。该模型的先进能力是通过一个精...",
      "extracted_excerpt": "本文介绍了 LongCat-Flash-Thinking，一个高效的、拥有 5600 亿参数的开源混合专家（MoE）推理模型。该模型的先进能力是通过一个精...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2024/03/01/gemini-15-unlocking-multimodal-understanding-across-millions-of-tokens-of-context.html",
      "title": "Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context",
      "raw_excerpt": "本文介绍了Gemini 1.5模型系列，这是一个新一代的高计算效率多模态模型家族。其核心研究目标是突破现有大语言模型（LLM）在上下文长度上的限制，实现对...",
      "extracted_excerpt": "本文介绍了Gemini 1.5模型系列，这是一个新一代的高计算效率多模态模型家族。其核心研究目标是突破现有大语言模型（LLM）在上下文长度上的限制，实现对...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/05/01/mimo-unlocking-the-reasoning-potential-of-language-model-from-pretraining-to-posttraining.html",
      "title": "MiMo: Unlocking the Reasoning Potential of Language Model – From Pretraining to Posttraining",
      "raw_excerpt": "本文介绍了为推理任务而生的 7B 参数大规模语言模型 MiMo-7B，其在预训练和后训练阶段都进行了优化。研究的核心问题是如何通过预训练和后训练策略的协同...",
      "extracted_excerpt": "本文介绍了为推理任务而生的 7B 参数大规模语言模型 MiMo-7B，其在预训练和后训练阶段都进行了优化。研究的核心问题是如何通过预训练和后训练策略的协同...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/05/01/qwen3-technical-report.html",
      "title": "Qwen3 Technical Report",
      "raw_excerpt": "本文介绍了Qwen模型家族的最新版本——Qwen3。Qwen3系列大型语言模型（LLMs）旨在提升性能、效率和多语言能力，其核心研究目标是通过一系列创新设...",
      "extracted_excerpt": "本文介绍了Qwen模型家族的最新版本——Qwen3。Qwen3系列大型语言模型（LLMs）旨在提升性能、效率和多语言能力，其核心研究目标是通过一系列创新设...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2023/02/01/llama-open-and-efficient-foundation-language-models.html",
      "title": "LLaMA: Open and Efficient Foundation Language Models",
      "raw_excerpt": "本文介绍了一个名为LLaMA（Large Language Model Meta AI）的基础语言模型集合，其参数规模从70亿（7B）到650亿（65B）...",
      "extracted_excerpt": "本文介绍了一个名为LLaMA（Large Language Model Meta AI）的基础语言模型集合，其参数规模从70亿（7B）到650亿（65B）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/08/01/kimi-k2-open-agentic-intelligence.html",
      "title": "KIMI K2: OPEN AGENTIC INTELLIGENCE",
      "raw_excerpt": "本文介绍了Kimi K2，一个拥有1.04万亿总参数和320亿激活参数的混合专家（MoE）大型语言模型，其设计旨在应对智能体能力的核心挑战并推动其边界。研...",
      "extracted_excerpt": "本文介绍了Kimi K2，一个拥有1.04万亿总参数和320亿激活参数的混合专家（MoE）大型语言模型，其设计旨在应对智能体能力的核心挑战并推动其边界。研...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/04/01/seed15-thinking-advancing-superb-reasoning-models-with-reinforcement-learning.html",
      "title": "Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning",
      "raw_excerpt": "本文介绍了一款名为Seed1.5-Thinking的新型推理模型，该模型通过在响应前进行思考来提升在广泛基准测试中的性能。研究的核心目标是开发一个在推理和...",
      "extracted_excerpt": "本文介绍了一款名为Seed1.5-Thinking的新型推理模型，该模型通过在响应前进行思考来提升在广泛基准测试中的性能。研究的核心目标是开发一个在推理和...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/08/01/longcat-flash-technical-report.html",
      "title": "LongCat-Flash Technical Report",
      "raw_excerpt": "LongCat-Flash 旨在沿两个协同方向推进语言模型的前沿：计算效率和智能体能力。本文的贡献涵盖了效率和智能体智能两个方面：",
      "extracted_excerpt": "LongCat-Flash 旨在沿两个协同方向推进语言模型的前沿：计算效率和智能体能力。本文的贡献涵盖了效率和智能体智能两个方面：",
      "excerpt_length": 65,
      "extracted_length": 65,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2024/01/01/deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.html",
      "title": "DeepSeek-Coder: When the Large Language Model Meets Programming - The Rise of Code Intelligence",
      "raw_excerpt": "本文旨在解决开源代码模型与闭源模型之间的性能差距。尽管大型语言模型（LLMs）极大地改变了软件开发中的代码智能，但强大的闭源模型由于其专有性，限制了广泛的...",
      "extracted_excerpt": "本文旨在解决开源代码模型与闭源模型之间的性能差距。尽管大型语言模型（LLMs）极大地改变了软件开发中的代码智能，但强大的闭源模型由于其专有性，限制了广泛的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2023/07/01/llama-2-open-foundation-and-fine-tuned-chat-models.html",
      "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models",
      "raw_excerpt": "本文旨在解决现有开源大语言模型（LLMs）虽在预训练性能上可与闭源模型媲美，但因缺乏与人类偏好对齐的深度微调，而无法成为ChatGPT、BARD等成熟“产...",
      "extracted_excerpt": "本文旨在解决现有开源大语言模型（LLMs）虽在预训练性能上可与闭源模型媲美，但因缺乏与人类偏好对齐的深度微调，而无法成为ChatGPT、BARD等成熟“产...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2024/12/01/deepseek-v3-technical-report.html",
      "title": "DeepSeek-V3 Technical Report",
      "raw_excerpt": "DeepSeek-V3是一个拥有6710亿总参数的强混合专家（MoE）语言模型，每个token激活370亿参数。为了实现经济高效的训练和推理，该模型继承并...",
      "extracted_excerpt": "DeepSeek-V3是一个拥有6710亿总参数的强混合专家（MoE）语言模型，每个token激活370亿参数。为了实现经济高效的训练和推理，该模型继承并...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2025/08/01/glm-45-agentic-reasoning-and-coding-arc-foundation-models.html",
      "title": "GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models",
      "raw_excerpt": "本文介绍了两个新的大型语言模型：GLM-4.5 和 GLM-4.5-Air，旨在统一代理（Agentic）、推理（Reasoning）和编码（Coding...",
      "extracted_excerpt": "本文介绍了两个新的大型语言模型：GLM-4.5 和 GLM-4.5-Air，旨在统一代理（Agentic）、推理（Reasoning）和编码（Coding...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/algorithm/models/2024/07/01/the-llama-3-herd-of-models.html",
      "title": "The Llama 3 Herd of Models",
      "raw_excerpt": "本文介绍了一套名为 Llama 3 的新型基础模型。Llama 3 模型家族原生支持多语言、编码、推理和工具使用。其中最大的模型是一个拥有4050亿（40...",
      "extracted_excerpt": "本文介绍了一套名为 Llama 3 的新型基础模型。Llama 3 模型家族原生支持多语言、编码、推理和工具使用。其中最大的模型是一个拥有4050亿（40...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2025/05/01/sageattention3-microscaling-fp4-attention-for-inference-and-an-exploration-of-8-bit-training.html",
      "title": "SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-bit Training",
      "raw_excerpt": "研究动机: Attention机制的效率对于生成模型至关重要，特别是其二次时间复杂度在处理长序列时成为瓶颈。量化是利用GPU中低比特张量核心（Tensor...",
      "extracted_excerpt": "研究动机: Attention机制的效率对于生成模型至关重要，特别是其二次时间复杂度在处理长序列时成为瓶颈。量化是利用GPU中低比特张量核心（Tensor...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2023/09/01/deepspeed-ulysses-system-optimizations-for-enabling-training-of-extreme-long-sequence-transformer-models.html",
      "title": "DEEPSPEED ULYSSES: SYSTEM OPTIMIZATIONS FOR ENABLING TRAINING OF EXTREME LONG SEQUENCE TRANSFORMER MODELS",
      "raw_excerpt": "本文针对大规模语言模型（LLM）在长序列训练中面临的系统挑战，介绍了一种名为 DeepSpeed-Ulysses 的新方法。",
      "extracted_excerpt": "本文针对大规模语言模型（LLM）在长序列训练中面临的系统挑战，介绍了一种名为 DeepSpeed-Ulysses 的新方法。",
      "excerpt_length": 62,
      "extracted_length": 62,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2019/11/01/fast-transformer-decoding-one-write-head-is-all-you-need.html",
      "title": "Fast Transformer Decoding: One Write-Head is All You Need",
      "raw_excerpt": "核心问题与研究目标：Transformer 神经网络序列模型在增量推理（incremental inference）时的速度是一个主要挑战。在现代计算硬件...",
      "extracted_excerpt": "核心问题与研究目标：Transformer 神经网络序列模型在增量推理（incremental inference）时的速度是一个主要挑战。在现代计算硬件...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2024/04/01/leave-no-context-behind-efficient-infinite-context-transformers-with-infini-attention.html",
      "title": "Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention",
      "raw_excerpt": "本文旨在解决 Transformer 及基于 Transformer 的大型语言模型（LLMs）因注意力机制的二次复杂度而在处理长序列时面临的内存和计算瓶...",
      "extracted_excerpt": "本文旨在解决 Transformer 及基于 Transformer 的大型语言模型（LLMs）因注意力机制的二次复杂度而在处理长序列时面临的内存和计算瓶...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2020/02/01/low-rank-bottleneck-in-multi-head-attention-models.html",
      "title": "Low-Rank Bottleneck in Multi-head Attention Models",
      "raw_excerpt": "本文旨在识别并解决当前多头注意力模型中一个关键的性能瓶颈。",
      "extracted_excerpt": "本文旨在识别并解决当前多头注意力模型中一个关键的性能瓶颈。",
      "excerpt_length": 29,
      "extracted_length": 29,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2023/10/01/ring-attention-with-blockwise-transformers-for-near-infinite-context.html",
      "title": "Ring Attention with Blockwise Transformers for Near-Infinite Context",
      "raw_excerpt": "核心问题: Transformer架构在处理长序列时面临严峻的内存挑战。其自注意力机制的内存成本与序列长度成二次方关系，导致难以扩展到长序列输入。此外，即...",
      "extracted_excerpt": "核心问题: Transformer架构在处理长序列时面临严峻的内存挑战。其自注意力机制的内存成本与序列长度成二次方关系，导致难以扩展到长序列输入。此外，即...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2025/09/01/sla-beyond-sparsity-in-diffusion-transformers-via-fine-tunable-sparselinear-attention.html",
      "title": "SLA: BEYOND SPARSITY IN DIFFUSION TRANSFORMERS VIA FINE-TUNABLE SPARSE–LINEAR ATTENTION",
      "raw_excerpt": "核心问题: 在扩散 Transformer（DiT）模型中，尤其是在视频生成领域，由于序列长度很长，注意力机制的二次方复杂度成为主要的延迟瓶颈。",
      "extracted_excerpt": "核心问题: 在扩散 Transformer（DiT）模型中，尤其是在视频生成领域，由于序列长度很长，注意力机制的二次方复杂度成为主要的延迟瓶颈。",
      "excerpt_length": 72,
      "extracted_length": 72,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2023/07/01/flashattention-2-faster-attention-with-better-parallelism-and-work-partitioning.html",
      "title": "FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning",
      "raw_excerpt": "本文的核心问题是解决 Transformer 模型在处理长序列时的性能瓶颈。注意力层的运行时间和内存占用随序列长度呈二次方增长，这限制了模型处理长文档、高...",
      "extracted_excerpt": "本文的核心问题是解决 Transformer 模型在处理长序列时的性能瓶颈。注意力层的运行时间和内存占用随序列长度呈二次方增长，这限制了模型处理长文档、高...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2025/02/01/tree-attention-topology-aware-decoding-for-long-context-attention-on-gpu-clusters.html",
      "title": "TREE ATTENTION: TOPOLOGY-AWARE DECODING FOR LONG-CONTEXT ATTENTION ON GPU CLUSTERS",
      "raw_excerpt": "本文旨在解决现代 Transformer 架构核心操作——自注意力机制的计算瓶颈问题。自注意力机制的计算复杂度随序列长度呈二次方增长，这使得处理长上下文的...",
      "extracted_excerpt": "本文旨在解决现代 Transformer 架构核心操作——自注意力机制的计算瓶颈问题。自注意力机制的计算复杂度随序列长度呈二次方增长，这使得处理长上下文的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2025/05/01/flashmla-etap-efficient-transpose-attention-pipeline-for-accelerating-mla-inference-on-nvidia-h20-gpus.html",
      "title": "FlashMLA-ETAP: Efficient Transpose Attention Pipeline for Accelerating MLA Inference on NVIDIA H20 GPUs",
      "raw_excerpt": "核心问题: Transformer架构中的注意力机制（如多头注意力MHA和多头潜在注意力MLA）具有与序列长度平方成正比的计算复杂度，这在处理长上下文任务...",
      "extracted_excerpt": "核心问题: Transformer架构中的注意力机制（如多头注意力MHA和多头潜在注意力MLA）具有与序列长度平方成正比的计算复杂度，这在处理长上下文任务...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2024/07/01/flashattention-3-fast-and-accurate-attention-with-asynchrony-and-low-precision.html",
      "title": "FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision",
      "raw_excerpt": "本文的核心问题是，尽管FlashAttention-2通过最小化内存读写加速了GPU上的注意力计算，但在最新的H100 GPU上其利用率仅达到35%，未能...",
      "extracted_excerpt": "本文的核心问题是，尽管FlashAttention-2通过最小化内存读写加速了GPU上的注意力计算，但在最新的H100 GPU上其利用率仅达到35%，未能...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2024/11/01/sageattention2-efficient-attention-with-thorough-outlier-smoothing-and-per-thread-int4-quantization.html",
      "title": "SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization",
      "raw_excerpt": "本文针对现有注意力计算加速方法的不足，提出了 SageAttention2，旨在通过使用速度更快的4位矩阵乘法（INT4 Matmul）并结合多种精度增强...",
      "extracted_excerpt": "本文针对现有注意力计算加速方法的不足，提出了 SageAttention2，旨在通过使用速度更快的4位矩阵乘法（INT4 Matmul）并结合多种精度增强...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2024/07/01/usp-a-unified-sequence-parallelism-approach-for-long-context-generative-ai.html",
      "title": "USP: A Unified Sequence Parallelism Approach for Long Context Generative AI",
      "raw_excerpt": "本文旨在解决生成式AI模型中日益增长的长上下文处理需求所带来的挑战。随着Claude、GPT-4、Gemini 1.5 Pro和Sora等模型将上下文长度...",
      "extracted_excerpt": "本文旨在解决生成式AI模型中日益增长的长上下文处理需求所带来的挑战。随着Claude、GPT-4、Gemini 1.5 Pro和Sora等模型将上下文长度...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2023/12/01/gqa-training-generalized-multi-query-transformer-models-from-multi-head-checkpoints.html",
      "title": "GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints",
      "raw_excerpt": "自回归解码器的推理过程是Transformer模型的一个严重瓶颈，主要因为在每个解码步骤都需要加载解码器权重以及所有的注意力键（keys）和值（value...",
      "extracted_excerpt": "自回归解码器的推理过程是Transformer模型的一个严重瓶颈，主要因为在每个解码步骤都需要加载解码器权重以及所有的注意力键（keys）和值（value...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2024/10/01/sageattention-accurate-8-bit-attention-for-plug-and-play-inference-acceleration.html",
      "title": "SAGEATTENTION: ACCURATE 8-BIT ATTENTION FOR PLUG-AND-PLAY INFERENCE ACCELERATION",
      "raw_excerpt": "本文旨在解决Transformer模型中注意力（Attention）机制的计算效率问题。随着序列长度的增加，具有 $O(N^2)$ 计算复杂度的注意力机制...",
      "extracted_excerpt": "本文旨在解决Transformer模型中注意力（Attention）机制的计算效率问题。随着序列长度的增加，具有 $O(N^2)$ 计算复杂度的注意力机制...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2025/08/01/mixture-of-contexts-for-long-video-generation.html",
      "title": "Mixture of Contexts for Long Video Generation",
      "raw_excerpt": "核心问题：长视频生成本质上是一个长上下文记忆问题。模型必须在长时间范围内保留和检索显著事件，而不会出现内容崩塌或漂移。然而，将扩散变换器（Diffusio...",
      "extracted_excerpt": "核心问题：长视频生成本质上是一个长上下文记忆问题。模型必须在长时间范围内保留和检索显著事件，而不会出现内容崩塌或漂移。然而，将扩散变换器（Diffusio...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/attention/2022/07/01/flashattention-fast-and-memory-efficient-exact-attention-with-io-awareness.html",
      "title": "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness",
      "raw_excerpt": "本文旨在解决Transformer模型在处理长序列时速度慢、内存消耗大的核心问题，该问题源于自注意力机制的时间和内存复杂度与序列长度成二次方关系。许多现有...",
      "extracted_excerpt": "本文旨在解决Transformer模型在处理长序列时速度慢、内存消耗大的核心问题，该问题源于自注意力机制的时间和内存复杂度与序列长度成二次方关系。许多现有...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2023/04/01/pytorch-fsdp-experiences-on-scaling-fully-sharded-data-parallel.html",
      "title": "PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel",
      "raw_excerpt": "本文介绍了PyTorch的完全分片数据并行（Fully Sharded Data Parallel, FSDP），这是一个用于大规模模型训练的工业级解决方...",
      "extracted_excerpt": "本文介绍了PyTorch的完全分片数据并行（Fully Sharded Data Parallel, FSDP），这是一个用于大规模模型训练的工业级解决方...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/08/01/optimus-accelerating-large-scale-multi-modal-llm-training-by-bubble-exploitation.html",
      "title": "Optimus: Accelerating Large-Scale Multi-Modal LLM Training by Bubble Exploitation",
      "raw_excerpt": "本文旨在解决训练大规模多模态语言模型（MLLM）时的效率低下问题。现有系统在训练MLLM时，由于异构的模态模型和3D并行中复杂的数据依赖性，会产生大量的G...",
      "extracted_excerpt": "本文旨在解决训练大规模多模态语言模型（MLLM）时的效率低下问题。现有系统在训练MLLM时，由于异构的模态模型和3D并行中复杂的数据依赖性，会产生大量的G...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/09/01/domino-eliminating-communication-in-llm-training-via-generic-tensor-slicing-and-overlapping.html",
      "title": "Domino: Eliminating Communication in LLM Training via Generic Tensor Slicing and Overlapping",
      "raw_excerpt": "生成式AI的最新进展在各种领域启用了新的应用场景，例如聊天机器人【[53], OpenAI. ChatGPT. https://chatgpt.com/,...",
      "extracted_excerpt": "生成式AI的最新进展在各种领域启用了新的应用场景，例如聊天机器人【[53], OpenAI. ChatGPT. https://chatgpt.com/,...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2021/07/01/chimera-efficiently-training-large-scale-neural-networks-with-bidirectional-pipelines.html",
      "title": "Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines",
      "raw_excerpt": "核心问题：训练大规模深度学习模型极具挑战性。随着模型规模的增长（例如，GPT-3拥有1750亿参数），简单的并行化方案（如数据并行）已不再适用，因为模型无...",
      "extracted_excerpt": "核心问题：训练大规模深度学习模型极具挑战性。随着模型规模的增长（例如，GPT-3拥有1750亿参数），简单的并行化方案（如数据并行）已不再适用，因为模型无...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/06/01/universal-checkpointing-a-flexible-and-efficient-distributed-checkpointing-system-for-large-scale-dnn-training-with-reconfigurable-parallelism.html",
      "title": "Universal Checkpointing: A Flexible and Efficient Distributed Checkpointing System for Large-Scale DNN Training with Reconfigurable Parallelism",
      "raw_excerpt": "本文旨在解决大规模深度神经网络（DNN）训练中的一个核心问题：现有训练系统在通过检查点技术重新配置并行策略方面的支持非常有限。随着模型规模、数据量和序列长...",
      "extracted_excerpt": "本文旨在解决大规模深度神经网络（DNN）训练中的一个核心问题：现有训练系统在通过检查点技术重新配置并行策略方面的支持非常有限。随着模型规模、数据量和序列长...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2023/10/01/fault-tolerant-hybrid-parallel-training-at-scale-with-reliable-and-efficient-in-memory-checkpointing.html",
      "title": "Fault-Tolerant Hybrid-Parallel Training at Scale with Reliable and Efficient In-memory Checkpointing",
      "raw_excerpt": "核心问题：为了高效扩展大型模型（LM）的训练，研究人员从数据并行（DP）转向GPU集群上的混合并行（HP）。然而，这些集群频繁遭遇硬件和软件故障。现有的内...",
      "extracted_excerpt": "核心问题：为了高效扩展大型模型（LM）的训练，研究人员从数据并行（DP）转向GPU集群上的混合并行（HP）。然而，这些集群频繁遭遇硬件和软件故障。现有的内...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/03/01/numerical-error-analysis-of-large-language-models.html",
      "title": "Numerical Error Analysis of Large Language Models",
      "raw_excerpt": "本文旨在对大型语言模型（LLMs）中有限数值精度计算导致的舍入误差进行理论分析，以揭示其对训练不稳定性的潜在影响。",
      "extracted_excerpt": "本文旨在对大型语言模型（LLMs）中有限数值精度计算导致的舍入误差进行理论分析，以揭示其对训练不稳定性的潜在影响。",
      "excerpt_length": 57,
      "extracted_length": 57,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/04/01/simai-unifying-architecture-design-and-performance-tuning-for-large-scale-large-language-model-training-with-scalability-and-precision.html",
      "title": "SimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large Language Model Training with Scalability and Precision",
      "raw_excerpt": "本文介绍了一个名为 SimAI 的统一模拟器，旨在精确高效地模拟大规模语言模型（LLM）的训练过程，以解决因 LLM 训练所需 GPU 数量庞大而导致新设...",
      "extracted_excerpt": "本文介绍了一个名为 SimAI 的统一模拟器，旨在精确高效地模拟大规模语言模型（LLM）的训练过程，以解决因 LLM 训练所需 GPU 数量庞大而导致新设...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/07/01/efficient-training-of-large-language-models-on-distributed-infrastructures-a-survey.html",
      "title": "Efficient Training of Large Language Models on Distributed Infrastructures: A Survey",
      "raw_excerpt": "本文旨在全面综述大型语言模型（LLM）训练系统和基础设施的最新进展，以应对其在“可扩展性、效率和可靠性”（SER）方面带来的重大挑战。\n*   核心问题:...",
      "extracted_excerpt": "本文旨在全面综述大型语言模型（LLM）训练系统和基础设施的最新进展，以应对其在“可扩展性、效率和可靠性”（SER）方面带来的重大挑战。\n*   核心问题:...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/01/01/obscura-concealing-recomputation-overhead-in-training-of-large-language-models-with-bubble-filling-pipeline-transformation.html",
      "title": "Obscura: Concealing Recomputation Overhead in Training of Large Language Models with Bubble-filling Pipeline Transformation",
      "raw_excerpt": "本文旨在解决流水线并行训练大语言模型（LLM）时，早期阶段面临的内存瓶颈问题。虽然重计算（recomputation）可以缓解此问题，但会引入额外的计算开销。",
      "extracted_excerpt": "本文旨在解决流水线并行训练大语言模型（LLM）时，早期阶段面临的内存瓶颈问题。虽然重计算（recomputation）可以缓解此问题，但会引入额外的计算开销。",
      "excerpt_length": 79,
      "extracted_length": 79,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/01/01/flexpipe-maximizing-training-efficiency-for-transformer-based-models-with-variable-length-inputs.html",
      "title": "FlexPipe: Maximizing Training Efficiency for Transformer-based Models with Variable-Length Inputs",
      "raw_excerpt": "本文针对Transformer模型在处理可变长度输入时训练效率低下的问题，提出了一种从分布式系统角度出发的解决方案FlexPipe。现有方法主要集中在单次...",
      "extracted_excerpt": "本文针对Transformer模型在处理可变长度输入时训练效率低下的问题，提出了一种从分布式系统角度出发的解决方案FlexPipe。现有方法主要集中在单次...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/02/01/training-llms-with-mxfp4.html",
      "title": "Training LLMs with MXFP4",
      "raw_excerpt": "本文旨在解决使用如 MXFP4 等低精度数据类型加速大语言模型（LLM）训练时遇到的模型质量显著下降的问题。尽管 MXFP4 在支持的硬件上矩阵乘法（GE...",
      "extracted_excerpt": "本文旨在解决使用如 MXFP4 等低精度数据类型加速大语言模型（LLM）训练时遇到的模型质量显著下降的问题。尽管 MXFP4 在支持的硬件上矩阵乘法（GE...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2021/06/01/lora-low-rank-adaptation-of-large-language-models.html",
      "title": "LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS",
      "raw_excerpt": "本文针对大型预训练语言模型在适应下游任务时面临的挑战，提出了一种名为“低秩适应”（Low-Rank Adaptation, LoRA）的高效方法。",
      "extracted_excerpt": "本文针对大型预训练语言模型在适应下游任务时面临的挑战，提出了一种名为“低秩适应”（Low-Rank Adaptation, LoRA）的高效方法。",
      "excerpt_length": 73,
      "extracted_length": 73,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2022/11/01/on-optimizing-the-communication-of-model-parallelism.html",
      "title": "ON OPTIMIZING THE COMMUNICATION OF MODEL PARALLELISM",
      "raw_excerpt": "本文研究了一种在规模化模型并行深度学习（DL）中新颖且重要的通信模式，我们称之为跨设备网格重分片（cross-mesh resharding）。当模型并行...",
      "extracted_excerpt": "本文研究了一种在规模化模型并行深度学习（DL）中新颖且重要的通信模式，我们称之为跨设备网格重分片（cross-mesh resharding）。当模型并行...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/05/01/pipeline-parallelism-with-controllable-memory.html",
      "title": "Pipeline Parallelism with Controllable Memory",
      "raw_excerpt": "本文旨在解决现有流水线并行（Pipeline Parallelism, PP）调度策略中存在的两个主要问题：流水线气泡（pipeline bubbles）...",
      "extracted_excerpt": "本文旨在解决现有流水线并行（Pipeline Parallelism, PP）调度策略中存在的两个主要问题：流水线气泡（pipeline bubbles）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/06/01/scaling-llama-3-training-with-efficient-parallelism-strategies.html",
      "title": "Scaling Llama 3 Training with Efficient Parallelism Strategies",
      "raw_excerpt": "本文详细介绍了用于Llama 3预训练的并行技术的设计与实现。为了在数万个GPU上实现高效训练，Llama 3采用了一种四维（4D）并行技术的组合：完全分...",
      "extracted_excerpt": "本文详细介绍了用于Llama 3预训练的并行技术的设计与实现。为了在数万个GPU上实现高效训练，Llama 3采用了一种四维（4D）并行技术的组合：完全分...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/03/01/wlb-llm-workload-balanced-4d-parallelism-for-large-language-model-training.html",
      "title": "WLB-LLM: Workload-Balanced 4D Parallelism for Large Language Model Training",
      "raw_excerpt": "本文提出WLB-LLM，一种用于大型语言模型训练的工作负载平衡的4D并行策略。核心问题在于大型语言模型训练中存在的负载不均衡问题，导致GPU利用率低下。具...",
      "extracted_excerpt": "本文提出WLB-LLM，一种用于大型语言模型训练的工作负载平衡的4D并行策略。核心问题在于大型语言模型训练中存在的负载不均衡问题，导致GPU利用率低下。具...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/09/01/robust-llm-training-infrastructure-at-bytedance.html",
      "title": "Robust LLM Training Infrastructure at ByteDance",
      "raw_excerpt": "本文介绍了ByteRobust，一个为大规模语言模型（LLM）的稳健和稳定训练量身定制的大规模GPU基础设施管理系统。当前LLM的训练规模已达数万个GPU...",
      "extracted_excerpt": "本文介绍了ByteRobust，一个为大规模语言模型（LLM）的稳健和稳定训练量身定制的大规模GPU基础设施管理系统。当前LLM的训练规模已达数万个GPU...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2018/06/01/pipedream-fast-and-efficient-pipeline-parallel-dnn-training.html",
      "title": "PipeDream: Fast and Efficient Pipeline Parallel DNN Training",
      "raw_excerpt": "本文针对深度神经网络（DNN）训练中通信开销成为瓶颈的问题，提出了一种名为 PipeDream 的新型分布式训练系统。随着模型规模增大和GPU算力提升，传...",
      "extracted_excerpt": "本文针对深度神经网络（DNN）训练中通信开销成为瓶颈的问题，提出了一种名为 PipeDream 的新型分布式训练系统。随着模型规模增大和GPU算力提升，传...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2021/04/01/zero-infinity-breaking-the-gpu-memory-wall-for-extreme-scale-deep-learning.html",
      "title": "ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning",
      "raw_excerpt": "本文旨在解决大规模深度学习模型训练中日益严峻的“GPU内存墙”问题。过去三年，最大的密集模型参数量增长了1000倍以上，而GPU内存仅增长了5倍，这使得训...",
      "extracted_excerpt": "本文旨在解决大规模深度学习模型训练中日益严峻的“GPU内存墙”问题。过去三年，最大的密集模型参数量增长了1000倍以上，而GPU内存仅增长了5倍，这使得训...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2021/08/01/efficient-large-scale-language-model-training-on-gpu-clusters-using-megatron-lm.html",
      "title": "Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM",
      "raw_excerpt": "本文旨在解决大规模语言模型训练中的两大挑战：单个GPU内存容量有限无法容纳巨大模型，以及巨大的计算量导致训练时间过长。为应对这些挑战，本文研究了如何高效地...",
      "extracted_excerpt": "本文旨在解决大规模语言模型训练中的两大挑战：单个GPU内存容量有限无法容纳巨大模型，以及巨大的计算量导致训练时间过长。为应对这些挑战，本文研究了如何高效地...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2022/04/01/palm-scaling-language-modeling-with-pathways.html",
      "title": "PaLM: Scaling Language Modeling with Pathways",
      "raw_excerpt": "本文介绍了一个5400亿参数、密集激活的Transformer语言模型，名为Pathways Language Model（PaLM）。该模型旨在探索模型...",
      "extracted_excerpt": "本文介绍了一个5400亿参数、密集激活的Transformer语言模型，名为Pathways Language Model（PaLM）。该模型旨在探索模型...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2020/03/01/zero-memory-optimizations-toward-training-trillion-parameter-models.html",
      "title": "ZeRO: Memory Optimizations Toward Training Trillion Parameter Models",
      "raw_excerpt": "本文旨在解决训练参数量从数十亿到万亿级别的大型深度学习模型所面临的内存限制和效率挑战。现有并行化方案（如数据并行、模型并行）在内存效率、计算/通信效率和易...",
      "extracted_excerpt": "本文旨在解决训练参数量从数十亿到万亿级别的大型深度学习模型所面临的内存限制和效率挑战。现有并行化方案（如数据并行、模型并行）在内存效率、计算/通信效率和易...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2022/05/01/pathways-asynchronous-distributed-dataflow-for-ml.html",
      "title": "PATHWAYS: ASYNCHRONOUS DISTRIBUTED DATAFLOW FOR ML",
      "raw_excerpt": "本文介绍了一种名为PATHWAYS的新型大规模分布式机器学习系统。当前系统在模型、硬件和软件的共同演进中，存在过度专精于现有工作负载的风险。PATHWAY...",
      "extracted_excerpt": "本文介绍了一种名为PATHWAYS的新型大规模分布式机器学习系统。当前系统在模型、硬件和软件的共同演进中，存在过度专精于现有工作负载的风险。PATHWAY...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/01/01/jenga-enhancing-llm-long-context-fine-tuning-with-contextual-token-sparsity.html",
      "title": "Jenga: Enhancing LLM Long-Context Fine-tuning with Contextual Token Sparsity",
      "raw_excerpt": "随着对综合文档分析【13, 26】、扩展多轮对话【76】和复杂代码库处理【51, 54】需求的增长，具有更大上下文窗口的大语言模型（LLM）正成为人工智能...",
      "extracted_excerpt": "随着对综合文档分析【13, 26】、扩展多轮对话【76】和复杂代码库处理【51, 54】需求的增长，具有更大上下文窗口的大语言模型（LLM）正成为人工智能...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/11/01/minder-faulty-machine-detection-for-large-scale-distributed-model-training.html",
      "title": "Minder: Faulty Machine Detection for Large-scale Distributed Model Training",
      "raw_excerpt": "本文旨在解决大规模分布式模型训练中关键的故障机器检测问题。在生产环境中，训练任务平均每天可能遇到两次故障，导致任务暂停数小时，现有的人工排查方法耗时费力且...",
      "extracted_excerpt": "本文旨在解决大规模分布式模型训练中关键的故障机器检测问题。在生产环境中，训练任务平均每天可能遇到两次故障，导致任务暂停数小时，现有的人工排查方法耗时费力且...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2020/06/01/gshard-scaling-giant-models-with-conditional-computation-and-automatic-sharding.html",
      "title": "GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding",
      "raw_excerpt": "本文旨在解决扩展神经网络时面临的计算成本、编程易用性和并行设备实现效率等挑战。",
      "extracted_excerpt": "本文旨在解决扩展神经网络时面临的计算成本、编程易用性和并行设备实现效率等挑战。",
      "excerpt_length": 39,
      "extracted_length": 39,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/02/01/weipipe-weight-pipeline-parallelism-for-communication-effective-long-context-large-model-training.html",
      "title": "WeiPipe: Weight Pipeline Parallelism for Communication-Effective Long-Context Large Model Training",
      "raw_excerpt": "本文旨在解决传统流水线并行（Pipeline Parallelism, PP）在训练长上下文大语言模型（LLMs）时遇到的通信挑战。随着模型上下文长度的增...",
      "extracted_excerpt": "本文旨在解决传统流水线并行（Pipeline Parallelism, PP）在训练长上下文大语言模型（LLMs）时遇到的通信挑战。随着模型上下文长度的增...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2023/11/01/zero-bubble-pipeline-parallelism.html",
      "title": "ZERO BUBBLE PIPELINE PARALLELISM",
      "raw_excerpt": "本文针对大规模分布式训练中的管道并行（PP）效率问题，焦点在于消除管道气泡，这些气泡被视为同步训练语义下不可避免的。研究目标是开发一种调度策略，实现零管道...",
      "extracted_excerpt": "本文针对大规模分布式训练中的管道并行（PP）效率问题，焦点在于消除管道气泡，这些气泡被视为同步训练语义下不可避免的。研究目标是开发一种调度策略，实现零管道...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/03/01/understanding-stragglers-in-large-model-training-using-what-if-analysis.html",
      "title": "Understanding Stragglers in Large Model Training Using What-if Analysis",
      "raw_excerpt": "本文针对大型语言模型（LLM）训练中的落后者（straggler）问题进行全面研究，使用从字节跳动LLM训练集群收集的五个月跟踪数据。核心问题是：落后者在...",
      "extracted_excerpt": "本文针对大型语言模型（LLM）训练中的落后者（straggler）问题进行全面研究，使用从字节跳动LLM训练集群收集的五个月跟踪数据。核心问题是：落后者在...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/01/01/zen-empowering-distributed-training-with-sparsity-driven-data-synchronization.html",
      "title": "Zen: Empowering Distributed Training with Sparsity-driven Data Synchronization",
      "raw_excerpt": "论文针对分布式训练中的梯度同步问题，揭示了稀疏张量的高稀疏性特性，并探讨了其同步方案的设计空间。研究目标是桥接稀疏张量同步的性能差距，找出充分利用稀疏性的...",
      "extracted_excerpt": "论文针对分布式训练中的梯度同步问题，揭示了稀疏张量的高稀疏性特性，并探讨了其同步方案的设计空间。研究目标是桥接稀疏张量同步的性能差距，找出充分利用稀疏性的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2020/03/01/megatron-lm-training-multi-billion-parameter-language-models-using-model-parallelism.html",
      "title": "Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism",
      "raw_excerpt": "本文提出了一种简单高效的层内模型并行方法，以解决因内存限制而难以训练超大规模Transformer模型的核心问题。研究目标是实现能够训练数十亿参数Tran...",
      "extracted_excerpt": "本文提出了一种简单高效的层内模型并行方法，以解决因内存限制而难以训练超大规模Transformer模型的核心问题。研究目标是实现能够训练数十亿参数Tran...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/01/01/enabling-efficient-gpu-communication-over-multiple-nics-with-fuselink.html",
      "title": "Enabling Efficient GPU Communication over Multiple NICs with FuseLink",
      "raw_excerpt": "论文从引言中提取的核心问题是：在分布式机器学习任务中，GPU间通信效率对性能至关重要，但现有系统在处理不均衡通信流量时无法充分利用多个NIC，导致热点NI...",
      "extracted_excerpt": "论文从引言中提取的核心问题是：在分布式机器学习任务中，GPU间通信效率对性能至关重要，但现有系统在处理不均衡通信流量时无法充分利用多个NIC，导致热点NI...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/08/01/optimus-accelerating-large-scale-multi-modal-llm-training-by-bubble-exploitation.html",
      "title": "Optimus: Accelerating Large-Scale Multi-Modal LLM Training by Bubble Exploitation",
      "raw_excerpt": "本文针对多模态大语言模型（MLLM）训练中存在的效率低下问题，提出了一个名为 Optimus 的分布式训练系统，旨在通过利用训练过程中的 GPU 空闲时间...",
      "extracted_excerpt": "本文针对多模态大语言模型（MLLM）训练中存在的效率低下问题，提出了一个名为 Optimus 的分布式训练系统，旨在通过利用训练过程中的 GPU 空闲时间...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2022/05/01/fastermoe-modeling-and-optimizing-training-of-large-scale-dynamic-pre-trained-models.html",
      "title": "FasterMoE: Modeling and Optimizing Training of Large-Scale Dynamic Pre-Trained Models",
      "raw_excerpt": "本文旨在解决训练大规模Mixture-of-Expert (MoE)模型时面临的效率挑战，这些挑战主要源于动态专家选择和灵活的MoE结构。\n核心问题:\n1...",
      "extracted_excerpt": "本文旨在解决训练大规模Mixture-of-Expert (MoE)模型时面临的效率挑战，这些挑战主要源于动态专家选择和灵活的MoE结构。\n核心问题:\n1...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/04/01/bytecheckpoint-a-unified-checkpointing-system-for-large-foundation-model-development.html",
      "title": "ByteCheckpoint: A Unified Checkpointing System for Large Foundation Model Development",
      "raw_excerpt": "本文针对大型基础模型（LFMs）的开发过程中，检查点保存以保留训练状态至关重要，用于在各种故障或GPU资源和并行配置变化时恢复训练。此外，保存的检查点被分...",
      "extracted_excerpt": "本文针对大型基础模型（LFMs）的开发过程中，检查点保存以保留训练状态至关重要，用于在各种故障或GPU资源和并行配置变化时恢复训练。此外，保存的检查点被分...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2023/10/01/gemini-fast-failure-recovery-in-distributed-training-with-in-memory-checkpoints.html",
      "title": "Gemini: Fast Failure Recovery in Distributed Training with In-Memory Checkpoints",
      "raw_excerpt": "本文介绍了一个名为 Gemini 的分布式训练系统，旨在通过利用主机CPU内存的高聚合带宽进行检查点操作，从而为大规模模型训练实现快速的故障恢复。",
      "extracted_excerpt": "本文介绍了一个名为 Gemini 的分布式训练系统，旨在通过利用主机CPU内存的高聚合带宽进行检查点操作，从而为大规模模型训练实现快速的故障恢复。",
      "excerpt_length": 73,
      "extracted_length": 73,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2021/02/01/checkfreq-frequent-fine-grained-dnn-checkpointing.html",
      "title": "CheckFreq: Frequent, Fine-Grained DNN Checkpointing",
      "raw_excerpt": "深度神经网络（DNN）的训练是一项资源密集且耗时的任务。训练过程中，模型在 GPU 上进行计算以学习权重，这个过程会重复多个轮次（epoch）。学习到的权...",
      "extracted_excerpt": "深度神经网络（DNN）的训练是一项资源密集且耗时的任务。训练过程中，模型在 GPU 上进行计算以学习权重，这个过程会重复多个轮次（epoch）。学习到的权...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/02/01/megascale-scaling-large-language-model-training-to-more-than-10000-gpus.html",
      "title": "MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs",
      "raw_excerpt": "本文介绍了MegaScale的设计、实现和工程经验，这是一个用于在超过10,000个GPU规模上训练大语言模型（LLM）的生产系统。在如此大的规模下训练L...",
      "extracted_excerpt": "本文介绍了MegaScale的设计、实现和工程经验，这是一个用于在超过10,000个GPU规模上训练大语言模型（LLM）的生产系统。在如此大的规模下训练L...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/06/01/meshslice-efficient-2d-tensor-parallelism-for-distributed-dnn-training.html",
      "title": "MeshSlice: Efficient 2D Tensor Parallelism for Distributed DNN Training",
      "raw_excerpt": "本文旨在解决大规模DNN模型分布式训练中张量并行（TP）的通信瓶颈问题。现有的1D TP因通信成本高而可扩展性有限，而2D TP虽然能通过将矩阵分片到2D...",
      "extracted_excerpt": "本文旨在解决大规模DNN模型分布式训练中张量并行（TP）的通信瓶颈问题。现有的1D TP因通信成本高而可扩展性有限，而2D TP虽然能通过将矩阵分片到2D...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2024/08/01/disttrain-addressing-model-and-data-heterogeneity-with-disaggregated-training-for-multimodal-large-language-models.html",
      "title": "DistTrain: Addressing Model and Data Heterogeneity with Disaggregated Training for Multimodal Large Language Models",
      "raw_excerpt": "本文提出 DistTrain，一个高效且自适应的框架，旨在通过解决多模态异构性问题，革新多模态大语言模型（Multimodal LLMs）的训练方式。在大...",
      "extracted_excerpt": "本文提出 DistTrain，一个高效且自适应的框架，旨在通过解决多模态异构性问题，革新多模态大语言模型（Multimodal LLMs）的训练方式。在大...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2021/01/01/zero-offload-democratizing-billion-scale-model-training.html",
      "title": "ZeRO-Offload: Democratizing Billion-Scale Model Training",
      "raw_excerpt": "本文旨在解决大规模深度学习模型训练因需要复杂模型重构和昂贵GPU集群而仅限于少数研究者的核心问题。研究目标是通过开发一种新颖的异构深度学习训练技术 ZeR...",
      "extracted_excerpt": "本文旨在解决大规模深度学习模型训练因需要复杂模型重构和昂贵GPU集群而仅限于少数研究者的核心问题。研究目标是通过开发一种新颖的异构深度学习训练技术 ZeR...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/02/01/bytescale-efficient-scaling-of-llm-training-with-a-2048k-context-length-on-more-than-12000-gpus.html",
      "title": "ByteScale: Efficient Scaling of LLM Training with a 2048K Context Length on More Than 12,000 GPUs",
      "raw_excerpt": "本文针对大规模语言模型（LLM）长短序列混合训练场景，提出了一个名为 ByteScale 的高效、灵活且可扩展的训练框架。现有框架通常采用静态的通信网格（...",
      "extracted_excerpt": "本文针对大规模语言模型（LLM）长短序列混合训练场景，提出了一个名为 ByteScale 的高效、灵活且可扩展的训练框架。现有框架通常采用静态的通信网格（...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/01/01/accelerating-design-space-exploration-for-llm-training-systems-with-multi-experiment-parallel-simulation.html",
      "title": "Accelerating Design Space Exploration for LLM Training Systems with Multi-experiment Parallel Simulation",
      "raw_excerpt": "核心问题: 随着大型语言模型（LLM）的快速发展，GPU集群的规模日益扩大，达到了数万甚至数十万级别。这极大地扩展了LLM训练系统的设计空间，涵盖了并行策...",
      "extracted_excerpt": "核心问题: 随着大型语言模型（LLM）的快速发展，GPU集群的规模日益扩大，达到了数万甚至数十万级别。这极大地扩展了LLM训练系统的设计空间，涵盖了并行策...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2019/07/01/gpipe-easy-scaling-with-micro-batch-pipeline-parallelism.html",
      "title": "GPipe: Easy Scaling with Micro-Batch Pipeline Parallelism",
      "raw_excerpt": "本文的核心问题在于，扩大深度神经网络的容量是提升模型质量的有效方法，但这常常受限于单个加速器（如GPU或TPU）的内存，需要开发专用的算法或基础设施，而这...",
      "extracted_excerpt": "本文的核心问题在于，扩大深度神经网络的容量是提升模型质量的有效方法，但这常常受限于单个加速器（如GPU或TPU）的内存，需要开发专用的算法或基础设施，而这...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/train/2025/02/01/mario-near-zero-cost-activation-checkpointing-in-pipeline-parallelism.html",
      "title": "Mario: Near Zero-cost Activation Checkpointing in Pipeline Parallelism",
      "raw_excerpt": "本文提出了 Mario，一个流水线优化器，旨在将激活检查点（activation checkpointing）技术以近乎零开销的方式集成到现有的流水线并行...",
      "extracted_excerpt": "本文提出了 Mario，一个流水线优化器，旨在将激活检查点（activation checkpointing）技术以近乎零开销的方式集成到现有的流水线并行...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/07/01/mooncake-a-kvcache-centric-disaggregated-architecture-for-llm-serving.html",
      "title": "Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving",
      "raw_excerpt": "本文介绍了Mooncake，一个为领先的LLM服务Kimi提供支持的服务平台。",
      "extracted_excerpt": "本文介绍了Mooncake，一个为领先的LLM服务Kimi提供支持的服务平台。",
      "excerpt_length": 39,
      "extracted_length": 39,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/05/01/splitwise-efficient-generative-llm-inference-using-phase-splitting.html",
      "title": "Splitwise: Efficient Generative LLM Inference Using Phase Splitting",
      "raw_excerpt": "核心问题: 生成式大型语言模型（LLM）的推理任务通常在昂贵且耗电的GPU上运行，而全球范围内的GPU产能正面临紧缩。这些推理任务具有两个计算特性截然不同...",
      "extracted_excerpt": "核心问题: 生成式大型语言模型（LLM）的推理任务通常在昂贵且耗电的GPU上运行，而全球范围内的GPU产能正面临紧缩。这些推理任务具有两个计算特性截然不同...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2022/11/01/efficiently-scaling-transformer-inference.html",
      "title": "EFFICIENTLY SCALING TRANSFORMER INFERENCE",
      "raw_excerpt": "核心问题与研究目标：随着Transformer模型参数扩展到千亿（如【6, Brown et al., Language models are few-s...",
      "extracted_excerpt": "核心问题与研究目标：随着Transformer模型参数扩展到千亿（如【6, Brown et al., Language models are few-s...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2023/11/01/striped-attention-faster-ring-attention-for-causal-transformers.html",
      "title": "STRIPED ATTENTION: FASTER RING ATTENTION FOR CAUSAL TRANSFORMERS",
      "raw_excerpt": "核心问题与研究目标：\n为了满足Transformer模型【22, Attention is all you need, 2017, NIPS】对日益增长的...",
      "extracted_excerpt": "核心问题与研究目标：\n为了满足Transformer模型【22, Attention is all you need, 2017, NIPS】对日益增长的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/08/01/nanoflow-towards-optimal-large-language-model-serving-throughput.html",
      "title": "NanoFlow: Towards Optimal Large Language Model Serving Throughput",
      "raw_excerpt": "本文首先通过详细分析指出，尽管大型语言模型（LLM）服务包含内存密集型组件（如自注意力机制），但对于大多数常见工作负载和模型而言，端到端的LLM服务实际上...",
      "extracted_excerpt": "本文首先通过详细分析指出，尽管大型语言模型（LLM）服务包含内存密集型组件（如自注意力机制），但对于大多数常见工作负载和模型而言，端到端的LLM服务实际上...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/07/01/helix-parallelism-rethinking-sharding-strategies-for-interactive-multi-million-token-llm-decoding.html",
      "title": "Helix Parallelism: Rethinking Sharding Strategies for Interactive Multi-Million-Token LLM Decoding",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）在处理百万级Token上下文时，在严格的“词元到词元延迟”（Token-to-Token Latency, TTL）约束...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）在处理百万级Token上下文时，在严格的“词元到词元延迟”（Token-to-Token Latency, TTL）约束...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/05/01/insights-into-deepseek-v3-scaling-challenges-and-reflections-on-hardware-for-ai-architectures.html",
      "title": "Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures",
      "raw_excerpt": "本文探讨了大型语言模型（LLMs）在快速扩展过程中暴露出的当前硬件架构的关键限制，包括内存容量、计算效率和互连带宽的约束。DeepSeek-V3模型在20...",
      "extracted_excerpt": "本文探讨了大型语言模型（LLMs）在快速扩展过程中暴露出的当前硬件架构的关键限制，包括内存容量、计算效率和互连带宽的约束。DeepSeek-V3模型在20...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/05/01/preble-efficient-distributed-prompt-scheduling-for-llm-serving.html",
      "title": "Preble: Efficient Distributed Prompt Scheduling for LLM Serving",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）服务中，由长且部分共享的提示（prompt）带来的新挑战。",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）服务中，由长且部分共享的提示（prompt）带来的新挑战。",
      "excerpt_length": 46,
      "extracted_length": 46,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/01/01/weaver-efficient-multi-llm-serving-with-attention-offloading.html",
      "title": "Weaver: Efficient Multi-LLM Serving with Attention Offloading",
      "raw_excerpt": "本文旨在解决多大型语言模型（Multi-LLM）服务场景中存在的效率问题。在此类场景中，通常存在请求分布不均的现象：少数“热门”模型处理绝大多数请求，而大...",
      "extracted_excerpt": "本文旨在解决多大型语言模型（Multi-LLM）服务场景中存在的效率问题。在此类场景中，通常存在请求分布不均的现象：少数“热门”模型处理绝大多数请求，而大...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/06/01/a-survey-of-llm-inference-systems.html",
      "title": "A Survey of LLM Inference Systems",
      "raw_excerpt": "本文是一篇关于大型语言模型（LLM）推理系统的综述。随着ChatGPT等服务的普及，对能够支持高性能模型服务的专用LLM推理系统的需求日益增长。这些系统设...",
      "extracted_excerpt": "本文是一篇关于大型语言模型（LLM）推理系统的综述。随着ChatGPT等服务的普及，对能够支持高性能模型服务的专用LLM推理系统的需求日益增长。这些系统设...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/09/01/defeating-nondeterminism-in-llm-inference.html",
      "title": "Defeating Nondeterminism in LLM Inference",
      "raw_excerpt": "本文探讨并解决了大型语言模型（LLM）推理中普遍存在的不确定性问题。即使在理论上应为确定性的贪婪采样（temperature=0）模式下，LLM的输出在实...",
      "extracted_excerpt": "本文探讨并解决了大型语言模型（LLM）推理中普遍存在的不确定性问题。即使在理论上应为确定性的贪婪采样（temperature=0）模式下，LLM的输出在实...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/09/01/mnemosyne-parallelization-strategies-for-efficiently-serving-multi-million-context-length-llm-inference-requests-without-approximations.html",
      "title": "Mnemosyne: Parallelization Strategies for Efficiently Serving Multi-Million Context Length LLM Inference Requests Without Approximations",
      "raw_excerpt": "本文旨在解决大语言模型（LLMs）在处理百万级甚至千万级长上下文推理请求时面临的独特挑战。现有的长上下文训练技术（如环形注意力和条带注意力）无法很好地适应...",
      "extracted_excerpt": "本文旨在解决大语言模型（LLMs）在处理百万级甚至千万级长上下文推理请求时面临的独特挑战。现有的长上下文训练技术（如环形注意力和条带注意力）无法很好地适应...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/07/01/sglang-efficient-execution-of-structured-language-model-programs.html",
      "title": "SGLang: Efficient Execution of Structured Language Model Programs",
      "raw_excerpt": "本文针对日益复杂的语言模型（LLM）应用中存在的编程和执行效率低下问题，提出了SGLang，一个用于高效执行复杂语言模型程序的系统。",
      "extracted_excerpt": "本文针对日益复杂的语言模型（LLM）应用中存在的编程和执行效率低下问题，提出了SGLang，一个用于高效执行复杂语言模型程序的系统。",
      "excerpt_length": 66,
      "extracted_length": 66,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2022/07/01/orca-a-distributed-serving-system-for-transformer-based-generative-models.html",
      "title": "Orca: A Distributed Serving System for Transformer-Based Generative Models",
      "raw_excerpt": "本文旨在解决现有推理服务系统在处理基于 Transformer 的生成式模型（如 GPT-3）时遇到的性能瓶颈。这类模型的自回归特性要求为单个推理请求多次...",
      "extracted_excerpt": "本文旨在解决现有推理服务系统在处理基于 Transformer 的生成式模型（如 GPT-3）时遇到的性能瓶颈。这类模型的自回归特性要求为单个推理请求多次...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2023/12/01/superserve-fine-grained-inference-serving-for-unpredictable-workloads.html",
      "title": "SuperServe: Fine-Grained Inference Serving for Unpredictable Workloads",
      "raw_excerpt": "机器学习（ML）模型日益部署在数据中心和边缘生产应用的关键路径上，这要求ML推理服务系统能够在不可预测和突发性的请求到达率下提供服务。这些系统需要在应用的...",
      "extracted_excerpt": "机器学习（ML）模型日益部署在数据中心和边缘生产应用的关键路径上，这要求ML推理服务系统能够在不可预测和突发性的请求到达率下提供服务。这些系统需要在应用的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2023/10/01/flash-decoding-for-long-context-inference.html",
      "title": "Flash-Decoding for long-context inference",
      "raw_excerpt": "核心问题:\n大型语言模型（LLM）的推理（或称“解码”）成本高昂。解码过程是迭代的，每生成一个新词元（token）都需要模型进行一次前向传播。虽然可以通过...",
      "extracted_excerpt": "核心问题:\n大型语言模型（LLM）的推理（或称“解码”）成本高昂。解码过程是迭代的，每生成一个新词元（token）都需要模型进行一次前向传播。虽然可以通过...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2023/08/01/sarathi-efficient-llm-inference-by-piggybacking-decodes-with-chunked-prefills.html",
      "title": "SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）推理中的两大效率瓶颈：1）解码（decode）阶段由于每次只生成一个token，导致GPU计算利用率低下；2）在使用流水...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）推理中的两大效率瓶颈：1）解码（decode）阶段由于每次只生成一个token，导致GPU计算利用率低下；2）在使用流水...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/07/01/megascale-infer-serving-mixture-of-experts-at-scale-with-disaggregated-expert-parallelism.html",
      "title": "MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism",
      "raw_excerpt": "MegaScale-Infer 是一个高效且成本有效的系统，用于服务大规模混合专家（MoE）模型。核心问题是：MoE的稀疏激活架构在推理过程中将前馈网络（...",
      "extracted_excerpt": "MegaScale-Infer 是一个高效且成本有效的系统，用于服务大规模混合专家（MoE）模型。核心问题是：MoE的稀疏激活架构在推理过程中将前馈网络（...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/04/01/tilus-a-virtual-machine-for-arbitrary-low-precision-gpgpu-computation-in-llm-serving.html",
      "title": "Tilus: A Virtual Machine for Arbitrary Low-Precision GPGPU Computation in LLM Serving",
      "raw_excerpt": "核心问题：大型语言模型（LLM）服务对计算资源，特别是内存带宽和计算吞吐量，提出了巨大需求。低精度计算是提高效率的关键技术，但现有方法存在局限。用于生成低...",
      "extracted_excerpt": "核心问题：大型语言模型（LLM）服务对计算资源，特别是内存带宽和计算吞吐量，提出了巨大需求。低精度计算是提高效率的关键技术，但现有方法存在局限。用于生成低...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/01/01/distserve-disaggregating-prefill-and-decoding-for-goodput-optimized-large-language-model-serving.html",
      "title": "DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving",
      "raw_excerpt": "本文旨在解决现有大型语言模型（LLM）服务系统中预填充（Prefill）和解码（Decoding）两个计算阶段共存（colocation）所引发的性能问题。",
      "extracted_excerpt": "本文旨在解决现有大型语言模型（LLM）服务系统中预填充（Prefill）和解码（Decoding）两个计算阶段共存（colocation）所引发的性能问题。",
      "excerpt_length": 78,
      "extracted_length": 78,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/07/01/step-3-is-large-yet-affordable-model-system-co-design-for-cost-effective-decoding.html",
      "title": "Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding",
      "raw_excerpt": "本文介绍了Step-3的模型-系统协同设计，该设计专门为测试时扩展范式（test-time scaling paradigm）而工程化，主要优化目标是最小...",
      "extracted_excerpt": "本文介绍了Step-3的模型-系统协同设计，该设计专门为测试时扩展范式（test-time scaling paradigm）而工程化，主要优化目标是最小...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2019/10/01/transformers-state-of-the-art-natural-language-processing.html",
      "title": "Transformers: State-of-the-Art Natural Language Processing",
      "raw_excerpt": "本文介绍了 Transformers，一个旨在向更广泛的机器学习社区开放自然语言处理（NLP）领域最新进展的开源库。",
      "extracted_excerpt": "本文介绍了 Transformers，一个旨在向更广泛的机器学习社区开放自然语言处理（NLP）领域最新进展的开源库。",
      "excerpt_length": 58,
      "extracted_length": 58,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2023/09/01/efficient-memory-management-for-large-language-model-serving-with-pagedattention.html",
      "title": "Efficient Memory Management for Large Language Model Serving with PagedAttention",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）服务中的一个核心瓶颈：KV缓存（KV cache）的内存管理。现有系统由于将每个请求的KV缓存存储在连续的内存空间中，导...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）服务中的一个核心瓶颈：KV缓存（KV cache）的内存管理。现有系统由于将每个请求的KV缓存存储在连续的内存空间中，导...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/01/01/qfactory-accelerating-quantized-large-language-model-serving-with-qtile-graphs.html",
      "title": "QFactory: Accelerating Quantized Large Language Model Serving with Qtile Graphs",
      "raw_excerpt": "核心问题：现有的量化大语言模型（QLLM）推理系统通常依赖于一种静态的“即时执行”（eager execution）范式来处理反量化操作。这种方法在遇到量...",
      "extracted_excerpt": "核心问题：现有的量化大语言模型（QLLM）推理系统通常依赖于一种静态的“即时执行”（eager execution）范式来处理反量化操作。这种方法在遇到量...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2024/05/01/efficient-heterogeneous-large-language-model-decoding-with-model-attention-disaggregation.html",
      "title": "Efficient Heterogeneous Large Language Model Decoding with Model-Attention Disaggregation",
      "raw_excerpt": "本文针对基于Transformer的大语言模型(LLM)在实际服务中，由于昂贵的计算优化型加速器使用效率低下而带来的挑战，提出了一种创新的解决方案。现有研...",
      "extracted_excerpt": "本文针对基于Transformer的大语言模型(LLM)在实际服务中，由于昂贵的计算优化型加速器使用效率低下而带来的挑战，提出了一种创新的解决方案。现有研...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/2025/06/01/lia-a-single-gpu-llm-inference-acceleration-with-cooperative-amx-enabled-cpu-gpu-computation-and-cxl-offloading.html",
      "title": "LIA: A Single-GPU LLM Inference Acceleration with Cooperative AMX-Enabled CPU-GPU Computation and CXL Offloading",
      "raw_excerpt": "核心问题:\n大型语言模型（LLM）的参数规模持续增长，导致在单块GPU上进行推理变得不可行，因为模型参数和中间值（如KV缓存和激活值）的存储需求超过了高端...",
      "extracted_excerpt": "核心问题:\n大型语言模型（LLM）的参数规模持续增长，导致在单块GPU上进行推理变得不可行，因为模型参数和中间值（如KV缓存和激活值）的存储需求超过了高端...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2018/11/01/blockwise-parallel-decoding-for-deep-autoregressive-models.html",
      "title": "Blockwise Parallel Decoding for Deep Autoregressive Models",
      "raw_excerpt": "核心问题：\n深度自回归序列到序列模型，如Transformer和卷积序列到序列模型，虽然在训练时利用并行计算取得了显著的速度提升，但在推理（生成）时仍然是...",
      "extracted_excerpt": "核心问题：\n深度自回归序列到序列模型，如Transformer和卷积序列到序列模型，虽然在训练时利用并行计算取得了显著的速度提升，但在推理（生成）时仍然是...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2024/06/01/medusa-simple-llm-inference-acceleration-framework-with-multiple-decoding-heads.html",
      "title": "MEDUSA: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）自回归解码过程中的一个核心瓶颈：由于每个token的生成都依赖于前一个token的输出，这种顺序计算导致推理延迟较高。从...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）自回归解码过程中的一个核心瓶颈：由于每个token的生成都依赖于前一个token的输出，这种顺序计算导致推理延迟较高。从...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2024/08/01/magicdec-breaking-the-latency-throughput-tradeoff-for-long-context-generation-with-speculative-decoding.html",
      "title": "MAGICDEC: BREAKING THE LATENCY-THROUGHPUT TRADEOFF FOR LONG CONTEXT GENERATION WITH SPECULATIVE DECODING",
      "raw_excerpt": "本文旨在解决在长上下文应用（如交互式聊天机器人、文档分析和智能体工作流）中，同时实现大型语言模型（LLM）服务的低延迟和高吞吐量这一挑战。",
      "extracted_excerpt": "本文旨在解决在长上下文应用（如交互式聊天机器人、文档分析和智能体工作流）中，同时实现大型语言模型（LLM）服务的低延迟和高吞吐量这一挑战。",
      "excerpt_length": 69,
      "extracted_length": 69,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2024/06/01/eagle-2-faster-inference-of-language-models-with-dynamic-draft-trees.html",
      "title": "EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees",
      "raw_excerpt": "本文旨在解决现代大型语言模型（LLM）推理成本高昂且耗时的问题。投机采样（speculative sampling）是一种有效的解决方案，其通过快速生成草...",
      "extracted_excerpt": "本文旨在解决现代大型语言模型（LLM）推理成本高昂且耗时的问题。投机采样（speculative sampling）是一种有效的解决方案，其通过快速生成草...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2025/03/01/eagle-3-scaling-up-inference-acceleration-of-large-language-models-via-training-time-test.html",
      "title": "EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test",
      "raw_excerpt": "本文的核心问题是现代大语言模型（LLM）因其自回归特性而导致推理成本高昂且速度缓慢。现有的推测采样方法如EAGLE通过在特征层面进行自回归并重用目标模型的...",
      "extracted_excerpt": "本文的核心问题是现代大语言模型（LLM）因其自回归特性而导致推理成本高昂且速度缓慢。现有的推测采样方法如EAGLE通过在特征层面进行自回归并重用目标模型的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2024/01/01/eagle-speculative-sampling-requires-rethinking-feature-uncertainty.html",
      "title": "EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty",
      "raw_excerpt": "核心问题: 自回归解码是大型语言模型（LLMs）的通用标准，它逐个生成token，导致生成过程缓慢且成本高昂。基于推测性采样的方法通过将过程分为低成本的草...",
      "extracted_excerpt": "核心问题: 自回归解码是大型语言模型（LLMs）的通用标准，它逐个生成token，导致生成过程缓慢且成本高昂。基于推测性采样的方法通过将过程分为低成本的草...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2024/04/01/better-faster-large-language-models-via-multi-token-prediction.html",
      "title": "Better & Faster Large Language Models via Multi-token Prediction",
      "raw_excerpt": "本文研究的核心问题是，当前大型语言模型（LLMs）普遍采用的“下一令牌预测”（next-token prediction）训练范式在样本效率上存在不足，模...",
      "extracted_excerpt": "本文研究的核心问题是，当前大型语言模型（LLMs）普遍采用的“下一令牌预测”（next-token prediction）训练范式在样本效率上存在不足，模...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/speculative_decoding/2023/05/01/fast-inference-from-transformers-via-speculative-decoding.html",
      "title": "Fast Inference from Transformers via Speculative Decoding",
      "raw_excerpt": "本文旨在解决大型自回归模型（如Transformer）推理速度慢的问题。这类模型的解码过程是串行的，即解码K个token需要K次模型的串行运行。现有加速方...",
      "extracted_excerpt": "本文旨在解决大型自回归模型（如Transformer）推理速度慢的问题。这类模型的解码过程是串行的，即解码K个token需要K次模型的串行运行。现有加速方...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2024/08/01/marlin-mixed-precision-auto-regressive-parallel-inference-on-large-language-models.html",
      "title": "MARLIN: Mixed-Precision Auto-Regressive Parallel Inference on Large Language Models",
      "raw_excerpt": "本文研究了在多客户端并行（即批处理）场景下，如何通过权重定点化技术加速大型语言模型（LLM）的推理。现有技术在单用户推理时能通过减少内存移动获得显著加速，...",
      "extracted_excerpt": "本文研究了在多客户端并行（即批处理）场景下，如何通过权重定点化技术加速大型语言模型（LLM）的推理。现有技术在单用户推理时能通过减少内存移动获得显著加速，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2025/05/01/recipes-for-pre-training-llms-with-mxfp8.html",
      "title": "Recipes for Pre-training LLMs with MXFP8",
      "raw_excerpt": "本文旨在为使用 NVIDIA Blackwell GPU 中引入的微缩放（Microscaling, MX）格式，特别是 MXFP8，进行大语言模型（LL...",
      "extracted_excerpt": "本文旨在为使用 NVIDIA Blackwell GPU 中引入的微缩放（Microscaling, MX）格式，特别是 MXFP8，进行大语言模型（LL...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2023/04/01/with-shared-microexponents-a-little-shifting-goes-a-long-way.html",
      "title": "With Shared Microexponents, A Little Shifting Goes a Long Way",
      "raw_excerpt": "本文旨在解决大规模人工智能（AI）模型（特别是大型语言模型）带来的高计算和能源成本问题。量化技术（将FP32张量值转换为更低成本的数值标准）是降低成本的关...",
      "extracted_excerpt": "本文旨在解决大规模人工智能（AI）模型（特别是大型语言模型）带来的高计算和能源成本问题。量化技术（将FP32张量值转换为更低成本的数值标准）是降低成本的关...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2023/04/01/stable-and-low-precision-training-for-large-scale-vision-language-models.html",
      "title": "Stable and low-precision training for large-scale vision-language models",
      "raw_excerpt": "本文旨在解决大规模语言-视觉模型训练中的两个关键问题：训练速度和训练稳定性，并以对比语言-图像预训练（CLIP）模型为研究背景。",
      "extracted_excerpt": "本文旨在解决大规模语言-视觉模型训练中的两个关键问题：训练速度和训练稳定性，并以对比语言-图像预训练（CLIP）模型为研究背景。",
      "excerpt_length": 64,
      "extracted_length": 64,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2022/08/01/fp8-quantization-the-power-of-the-exponent.html",
      "title": "FP8 Quantization: The Power of the Exponent",
      "raw_excerpt": "本文深入研究了在神经网络推理中使用低比特浮点数（FP）格式的潜在优势，特别是与广泛使用的整数（INT）格式相比。研究的核心问题是，浮点格式通过将部分比特分...",
      "extracted_excerpt": "本文深入研究了在神经网络推理中使用低比特浮点数（FP）格式的潜在优势，特别是与广泛使用的整数（INT）格式相比。研究的核心问题是，浮点格式通过将部分比特分...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2024/11/01/microscopiq-accelerating-foundational-models-through-outlier-aware-microscaling-quantization.html",
      "title": "MicroScopiQ: Accelerating Foundational Models through Outlier-Aware Microscaling Quantization",
      "raw_excerpt": "本文旨在解决基础模型（FMs）量化中的一个核心挑战：离群点（outliers） 的存在。这些大幅值特征使得FMs的量化比传统DNNs更为困难。现有的离群点...",
      "extracted_excerpt": "本文旨在解决基础模型（FMs）量化中的一个核心挑战：离群点（outliers） 的存在。这些大幅值特征使得FMs的量化比传统DNNs更为困难。现有的离群点...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2023/09/01/ocp-microscaling-formats-mx-specification.html",
      "title": "OCP Microscaling Formats (MX) Specification",
      "raw_excerpt": "本规范旨在标准化一个开放、公开且可互操作的数据格式家族——微缩放（Microscaling, MX）格式，以应对人工智能（AI）训练和推理中日益增长的计算...",
      "extracted_excerpt": "本规范旨在标准化一个开放、公开且可互操作的数据格式家族——微缩放（Microscaling, MX）格式，以应对人工智能（AI）训练和推理中日益增长的计算...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2024/08/01/lut-tensor-core-a-software-hardware-co-design-for-lut-based-low-bit-llm-inference.html",
      "title": "LUT Tensor Core: A Software-Hardware Co-Design for LUT-Based Low-Bit LLM Inference",
      "raw_excerpt": "本文针对大规模语言模型（LLM）推理中资源密集的问题，提出了一种名为LUT Tensor Core的软硬件协同设计，旨在高效执行基于查找表（LUT）的混合...",
      "extracted_excerpt": "本文针对大规模语言模型（LLM）推理中资源密集的问题，提出了一种名为LUT Tensor Core的软硬件协同设计，旨在高效执行基于查找表（LUT）的混合...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2022/11/01/smoothquant-accurate-and-efficient-post-training-quantization-for-large-language-models.html",
      "title": "SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）因模型规模巨大而带来的计算和内存密集问题。现有量化方法在保持准确性和硬件效率方面难以兼顾：例如，ZeroQuant在处理...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）因模型规模巨大而带来的计算和内存密集问题。现有量化方法在保持准确性和硬件效率方面难以兼顾：例如，ZeroQuant在处理...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2025/03/01/oaken-fast-and-efficient-llm-serving-with-online-offline-hybrid-kv-cache-quantization.html",
      "title": "Oaken: Fast and Efficient LLM Serving with Online-Offline Hybrid KV Cache Quantization",
      "raw_excerpt": "大型语言模型（LLM）服务系统通过批处理多个请求以实现高吞吐量，但注意力操作的批处理极具挑战性，使得内存带宽成为关键瓶颈。当前，业界依赖配备多个高带宽内存...",
      "extracted_excerpt": "大型语言模型（LLM）服务系统通过批处理多个请求以实现高吞吐量，但注意力操作的批处理极具挑战性，使得内存带宽成为关键瓶颈。当前，业界依赖配备多个高带宽内存...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2023/06/01/awq-activation-aware-weight-quantization-for-on-device-llm-compression-and-acceleration.html",
      "title": "AWQ: ACTIVATION-AWARE WEIGHT QUANTIZATION FOR ON-DEVICE LLM COMPRESSION AND ACCELERATION",
      "raw_excerpt": "本文旨在解决在边缘设备上部署大语言模型（LLM）所面临的巨大模型尺寸和有限硬件资源的挑战。现有方案中，量化感知训练（QAT）成本过高，而训练后量化（PTQ...",
      "extracted_excerpt": "本文旨在解决在边缘设备上部署大语言模型（LLM）所面临的巨大模型尺寸和有限硬件资源的挑战。现有方案中，量化感知训练（QAT）成本过高，而训练后量化（PTQ...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2024/02/01/massive-activations-in-large-language-models.html",
      "title": "Massive Activations in Large Language Models",
      "raw_excerpt": "本文发现并研究了大型语言模型（LLMs）内部表征中一个令人惊讶的现象：极少数激活值表现出异常巨大的量级，我们称之为“巨量激活”（massive activ...",
      "extracted_excerpt": "本文发现并研究了大型语言模型（LLMs）内部表征中一个令人惊讶的现象：极少数激活值表现出异常巨大的量级，我们称之为“巨量激活”（massive activ...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2023/06/01/fp8-versus-int8-for-efficient-deep-learning-inference.html",
      "title": "FP8 versus INT8 for efficient deep learning inference",
      "raw_excerpt": "本文旨在论证，尽管FP8格式在神经网络训练领域备受关注，但它不太可能取代INT8成为边缘设备高效推理的主流格式。研究的核心问题是，在推理效率和准确性方面，...",
      "extracted_excerpt": "本文旨在论证，尽管FP8格式在神经网络训练领域备受关注，但它不太可能取代INT8成为边缘设备高效推理的主流格式。研究的核心问题是，在推理效率和准确性方面，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2024/01/01/fp6-llm-efficiently-serving-large-language-models-through-fp6-centric-algorithm-system-co-design.html",
      "title": "FP6-LLM: Efficiently Serving Large Language Models Through FP6-Centric Algorithm-System Co-Design",
      "raw_excerpt": "本文旨在解决大型语言模型（LLM）部署中的显存和内存带宽瓶颈问题，特别是针对6比特浮点（FP6）量化在现有GPU系统上缺乏高效支持的现状。尽管FP6在模型...",
      "extracted_excerpt": "本文旨在解决大型语言模型（LLM）部署中的显存和内存带宽瓶颈问题，特别是针对6比特浮点（FP6）量化在现有GPU系统上缺乏高效支持的现状。尽管FP6在模型...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2023/05/01/integer-or-floating-point-new-outlooks-for-low-bit-quantization-on-large-language-models.html",
      "title": "Integer or Floating Point? New Outlooks for Low-Bit Quantization on Large Language Models",
      "raw_excerpt": "本文探讨了在大型语言模型（LLMs）的低比特量化中，整数（INT）与浮点数（FP）格式的优劣问题。随着LLM规模的急剧增长，低比特量化对于降低模型大小和推...",
      "extracted_excerpt": "本文探讨了在大型语言模型（LLMs）的低比特量化中，整数（INT）与浮点数（FP）格式的优劣问题。随着LLM规模的急剧增长，低比特量化对于降低模型大小和推...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2025/01/01/decdec-a-systems-approach-to-advancing-low-bit-llm-quantization.html",
      "title": "DecDEC: A Systems Approach to Advancing Low-Bit LLM Quantization",
      "raw_excerpt": "本文针对大型语言模型（LLM）的量化问题，特别是低比特设置（如3比特和4比特精度）下模型质量下降的挑战，提出了一种推理方案DecDEC。该方案旨在改善低比...",
      "extracted_excerpt": "本文针对大型语言模型（LLM）的量化问题，特别是低比特设置（如3比特和4比特精度）下模型质量下降的挑战，提出了一种推理方案DecDEC。该方案旨在改善低比...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/low_precision/2022/09/01/fp8-formats-for-deep-learning.html",
      "title": "FP8 FORMATS FOR DEEP LEARNING",
      "raw_excerpt": "本文针对深度学习训练和推理加速的需求，提出了一种8位浮点数（FP8）的二进制交换格式。随着深度学习模型规模和训练所需计算资源的持续增长，例如GPT-3、T...",
      "extracted_excerpt": "本文针对深度学习训练和推理加速的需求，提出了一种8位浮点数（FP8）的二进制交换格式。随着深度学习模型规模和训练所需计算资源的持续增长，例如GPT-3、T...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2024/10/01/do-large-language-models-need-a-content-delivery-network.html",
      "title": "Do Large Language Models Need a Content Delivery Network?",
      "raw_excerpt": "随着大型语言模型（LLM）的快速发展，为各种LLM查询补充知识的需求也随之增长。因此，实现模块化、高效地在LLM推理中注入新知识至关重要。本文从系统架构的...",
      "extracted_excerpt": "随着大型语言模型（LLM）的快速发展，为各种LLM查询补充知识的需求也随之增长。因此，实现模块化、高效地在LLM推理中注入新知识至关重要。本文从系统架构的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2023/10/01/cachegen-kv-cache-compression-and-streaming-for-fast-large-language-model-serving.html",
      "title": "CacheGen: KV Cache Compression and Streaming for Fast Large Language Model Serving",
      "raw_excerpt": "图2：不同上下文加载方式如何影响网络延迟（传输上下文或KV缓存）和计算延迟（在上下文上运行注意力模块）。",
      "extracted_excerpt": "图2：不同上下文加载方式如何影响网络延迟（传输上下文或KV缓存）和计算延迟（在上下文上运行注意力模块）。",
      "excerpt_length": 52,
      "extracted_length": 52,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2024/07/01/cost-efficient-large-language-model-serving-for-multi-turn-conversations-with-cachedattention.html",
      "title": "Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention",
      "raw_excerpt": "本文旨在解决现有大型语言模型（LLM）服务引擎在执行多轮对话时效率低下的问题。核心问题在于，这些引擎需要重复计算历史token的键值（KV）缓存，从而导致...",
      "extracted_excerpt": "本文旨在解决现有大型语言模型（LLM）服务引擎在执行多轮对话时效率低下的问题。核心问题在于，这些引擎需要重复计算历史token的键值（KV）缓存，从而导致...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2024/04/01/prompt-cache-modular-attention-reuse-for-low-latency-inference.html",
      "title": "PROMPT CACHE: MODULAR ATTENTION REUSE FOR LOW-LATENCY INFERENCE",
      "raw_excerpt": "核心问题: 大量大型语言模型（LLM）的提示（prompt）包含频繁重用的文本片段，例如系统消息、提示模板以及用于上下文的文档。这种重叠导致在推理过程中存...",
      "extracted_excerpt": "核心问题: 大量大型语言模型（LLM）的提示（prompt）包含频繁重用的文本片段，例如系统消息、提示模板以及用于上下文的文档。这种重叠导致在推理过程中存...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2024/05/01/cacheblend-fast-large-language-model-serving-for-rag-with-cached-knowledge-fusion.html",
      "title": "CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion",
      "raw_excerpt": "本文旨在解决一个核心挑战：当一个大型语言模型（LLM）的输入包含多个文本块时，如何快速地组合它们各自预计算的键值缓存（KV Caches），以达到与昂贵的...",
      "extracted_excerpt": "本文旨在解决一个核心挑战：当一个大型语言模型（LLM）的输入包含多个文本块时，如何快速地组合它们各自预计算的键值缓存（KV Caches），以达到与昂贵的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2025/06/01/kvcache-cache-in-the-wild-characterizing-and-optimizing-kvcache-cache-at-a-large-cloud-provider.html",
      "title": "KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider",
      "raw_excerpt": "本文针对大型语言模型（LLM）服务中的一个关键性能优化技术——KV Cache（KV$）缓存，进行了首次系统性的生产环境工作负载特征分析。研究的核心问题在...",
      "extracted_excerpt": "本文针对大型语言模型（LLM）服务中的一个关键性能优化技术——KV Cache（KV$）缓存，进行了首次系统性的生产环境工作负载特征分析。研究的核心问题在...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/inference/kvcache/2025/05/01/ecco-improving-memory-bandwidth-and-capacity-for-llms-via-entropy-aware-cache-compression.html",
      "title": "Ecco: Improving Memory Bandwidth and Capacity for LLMs via Entropy-aware Cache Compression",
      "raw_excerpt": "大型语言模型（LLMs）在人工智能应用中展现了变革性能力，但其巨大的内存和计算需求，尤其是在资源受限的环境下，阻碍了其部署。量化技术作为一种关键解决方案，...",
      "extracted_excerpt": "大型语言模型（LLMs）在人工智能应用中展现了变革性能力，但其巨大的内存和计算需求，尤其是在资源受限的环境下，阻碍了其部署。量化技术作为一种关键解决方案，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/07/01/distflow-a-fully-distributed-rl-framework-for-scalable-and-efficient-llm-post-training.html",
      "title": "DistFlow: A Fully Distributed RL Framework for Scalable and Efficient LLM Post-Training",
      "raw_excerpt": "本文旨在解决现有大规模强化学习（RL）框架中的核心性能瓶颈问题。主流框架通常采用混合控制器架构，其中单控制器负责分派整体执行逻辑和管理数据传输，而多控制器...",
      "extracted_excerpt": "本文旨在解决现有大规模强化学习（RL）框架中的核心性能瓶颈问题。主流框架通常采用混合控制器架构，其中单控制器负责分派整体执行逻辑和管理数据传输，而多控制器...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2024/09/01/rlhfuse-efficient-rlhf-training-for-large-language-models-with-inter-and-intra-stage-fusion.html",
      "title": "RLHFuse: Efficient RLHF Training for Large Language Models with Inter- and Intra-Stage Fusion",
      "raw_excerpt": "本文旨在解决现有强化学习人类反馈（RLHF）训练系统在生产部署中GPU利用率低下的问题。核心问题源于RLHF训练的内在特性：生成阶段的数据倾斜（长尾效应）...",
      "extracted_excerpt": "本文旨在解决现有强化学习人类反馈（RLHF）训练系统在生产部署中GPU利用率低下的问题。核心问题源于RLHF训练的内在特性：生成阶段的数据倾斜（长尾效应）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2019/10/01/seed-rl-scalable-and-efficient-deep-rl-with-accelerated-central-inference.html",
      "title": "SEED RL: SCALABLE AND EFFICIENT DEEP-RL WITH ACCELERATED CENTRAL INFERENCE",
      "raw_excerpt": "本文旨在解决强化学习（RL）中的两大挑战：扩展性问题和现代计算资源（特别是加速器）的有效利用问题。当前先进的分布式算法在处理RL固有的异构任务（环境运行、...",
      "extracted_excerpt": "本文旨在解决强化学习（RL）中的两大挑战：扩展性问题和现代计算资源（特别是加速器）的有效利用问题。当前先进的分布式算法在处理RL固有的异构任务（环境运行、...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/08/01/seamlessflow-a-traineragent-isolation-rl-framework-achieving-bubble-free-pipelines-via-tag-scheduling.html",
      "title": "SeamlessFlow: A Trainer–Agent Isolation RL Framework Achieving Bubble-Free Pipelines via Tag Scheduling",
      "raw_excerpt": "本文旨在解决工业级强化学习（RL）部署中的两个核心挑战：(1) 将RL训练与复杂的智能体（agent）执行流解耦；(2) 在保持大规模部署所需的稳定性和可...",
      "extracted_excerpt": "本文旨在解决工业级强化学习（RL）部署中的两个核心挑战：(1) 将RL训练与复杂的智能体（agent）执行流解耦；(2) 在保持大规模部署所需的稳定性和可...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2022/06/01/gmi-drl-empowering-multi-gpu-deep-reinforcement-learning-with-gpu-spatial-multiplexing.html",
      "title": "GMI-DRL: Empowering Multi-GPU Deep Reinforcement Learning with GPU Spatial Multiplexing",
      "raw_excerpt": "核心问题: 深度强化学习（DRL）具有异构的计算工作负载（物理模拟 vs. GEMM-based神经网络操作）和交错执行的模式，这导致在现代强大的GPU平...",
      "extracted_excerpt": "核心问题: 深度强化学习（DRL）具有异构的计算工作负载（物理模拟 vs. GEMM-based神经网络操作）和交错执行的模式，这导致在现代强大的GPU平...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/05/01/areal-a-large-scale-asynchronous-reinforcement-learning-system-for-language-reasoning.html",
      "title": "AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning",
      "raw_excerpt": "强化学习（RL）已成为训练大型语言模型（LLM）的一种流行范式，尤其是在推理任务中。有效的LLM RL需要大规模并行化，并对高效训练系统提出迫切需求。大多...",
      "extracted_excerpt": "强化学习（RL）已成为训练大型语言模型（LLM）的一种流行范式，尤其是在推理任务中。有效的LLM RL需要大规模并行化，并对高效训练系统提出迫切需求。大多...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/04/01/retool-reinforcement-learning-for-strategic-tool-use-in-llms.html",
      "title": "ReTool: Reinforcement Learning for Strategic Tool Use in LLMs",
      "raw_excerpt": "本文旨在解决当前通过强化学习（RL）训练的推理模型（如DeepSeek R1）在纯文本推理方面表现出色，但在需要结构化问题解决的场景（如几何推理、精确计算...",
      "extracted_excerpt": "本文旨在解决当前通过强化学习（RL）训练的推理模型（如DeepSeek R1）在纯文本推理方面表现出色，但在需要结构化问题解决的场景（如几何推理、精确计算...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/04/01/streamrl-scalable-heterogeneous-and-elastic-rl-for-llms-with-disaggregated-stream-generation.html",
      "title": "StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation",
      "raw_excerpt": "本文探讨了大型语言模型（LLM）的强化学习（RL）训练架构。RL训练通常包括在线生成样本（generation）和使用这些样本进行训练（training）...",
      "extracted_excerpt": "本文探讨了大型语言模型（LLM）的强化学习（RL）训练架构。RL训练通常包括在线生成样本（generation）和使用这些样本进行训练（training）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2023/03/01/deepspeed-chat-easy-fast-and-affordable-rlhf-training-of-chatgpt-like-models-at-all-scales.html",
      "title": "DeepSpeed-Chat: Easy, Fast and Affordable RLHF Training of ChatGPT-like Models at All Scales",
      "raw_excerpt": "本文介绍了DeepSpeed-Chat系统，该系统旨在解决当前缺乏易用、高效且经济实惠的端到端RLHF（带有人类反馈的强化学习）训练管道的问题，特别是针对...",
      "extracted_excerpt": "本文介绍了DeepSpeed-Chat系统，该系统旨在解决当前缺乏易用、高效且经济实惠的端到端RLHF（带有人类反馈的强化学习）训练管道的问题，特别是针对...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/04/01/deepcoder-a-fully-open-source-14b-coder-at-o3-mini-level.html",
      "title": "DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level",
      "raw_excerpt": "本文由Agentica团队与Together AI合作，发布了DeepCoder-14B-Preview，这是一个通过分布式强化学习（RL）从Deepse...",
      "extracted_excerpt": "本文由Agentica团队与Together AI合作，发布了DeepCoder-14B-Preview，这是一个通过分布式强化学习（RL）从Deepse...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/08/01/rstar2-agent-agentic-reasoning-technical-report.html",
      "title": "rStar2-Agent: Agentic Reasoning Technical Report",
      "raw_excerpt": "本文介绍了 rStar2-Agent，一个通过智能体强化学习（agentic reinforcement learning）训练的 14B 数学推理模型，...",
      "extracted_excerpt": "本文介绍了 rStar2-Agent，一个通过智能体强化学习（agentic reinforcement learning）训练的 14B 数学推理模型，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/08/01/your-efficient-rl-framework-secretly-brings-you-offpolicy-rl-training.html",
      "title": "Your Efficient RL Framework Secretly Brings You OffPolicy RL Training",
      "raw_excerpt": "本文的核心研究问题在于，现代强化学习（RL）训练框架（如VeRL）为了提升效率，通常采用混合设计，即使用高度优化的推理引擎（如vLLM）进行rollout...",
      "extracted_excerpt": "本文的核心研究问题在于，现代强化学习（RL）训练框架（如VeRL）为了提升效率，通常采用混合设计，即使用高度优化的推理引擎（如vLLM）进行rollout...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2024/05/01/openrlhf-an-easy-to-use-scalable-and-high-performance-rlhf-framework.html",
      "title": "OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework",
      "raw_excerpt": "OpenRLHF 旨在解决现有强化学习从人类反馈（RLHF）和从可验证奖励中强化学习（RLVR）框架中普遍存在的推理瓶颈和复杂性障碍问题。研究目标是提供一...",
      "extracted_excerpt": "OpenRLHF 旨在解决现有强化学习从人类反馈（RLHF）和从可验证奖励中强化学习（RLVR）框架中普遍存在的推理瓶颈和复杂性障碍问题。研究目标是提供一...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/07/01/asyncflow-an-asynchronous-streaming-rl-framework-for-efficient-llm-post-training.html",
      "title": "AsyncFlow: An Asynchronous Streaming RL Framework for Efficient LLM Post-Training",
      "raw_excerpt": "本文针对大规模语言模型（LLM）后训练（Post-Training）阶段中强化学习（RL）框架面临的挑战，提出了一个名为AsyncFlow的异步流式RL框架。",
      "extracted_excerpt": "本文针对大规模语言模型（LLM）后训练（Post-Training）阶段中强化学习（RL）框架面临的挑战，提出了一个名为AsyncFlow的异步流式RL框架。",
      "excerpt_length": 79,
      "extracted_length": 79,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2024/10/01/hybridflow-a-flexible-and-efficient-rlhf-framework.html",
      "title": "HybridFlow: A Flexible and Efficient RLHF Framework",
      "raw_excerpt": "论文的核心问题是RLHF在LLM对齐中的复杂数据流表示和执行效率问题：传统RL框架使用单一控制器导致分布式计算开销大，现有的RLHF系统采用多控制器范式但...",
      "extracted_excerpt": "论文的核心问题是RLHF在LLM对齐中的复杂数据流表示和执行效率问题：传统RL框架使用单一控制器导致分布式计算开销大，现有的RLHF系统采用多控制器范式但...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/rl/2025/10/01/stabilizing-moe-reinforcement-learning-by-aligning-training-and-inference-routers.html",
      "title": "Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers",
      "raw_excerpt": "本文研究了在混合专家（MoE）模型中，强化学习（RL）训练的不稳定性问题。",
      "extracted_excerpt": "本文研究了在混合专家（MoE）模型中，强化学习（RL）训练的不稳定性问题。",
      "excerpt_length": 37,
      "extracted_length": 37,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2022/02/01/st-moe-designing-stable-and-transferable-sparse-expert-models.html",
      "title": "ST-MOE: DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS",
      "raw_excerpt": "本文旨在提高稀疏专家模型（Sparse Expert Models）的实用性和可靠性，解决其在训练稳定性和微调（fine-tuning）质量方面存在的问题...",
      "extracted_excerpt": "本文旨在提高稀疏专家模型（Sparse Expert Models）的实用性和可靠性，解决其在训练稳定性和微调（fine-tuning）质量方面存在的问题...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2023/03/01/scaling-vision-language-models-with-sparse-mixture-of-experts.html",
      "title": "Scaling Vision-Language Models with Sparse Mixture of Experts",
      "raw_excerpt": "本文探讨了使用稀疏门控专家混合（MoE）技术来扩展视觉-语言模型（VLM）的有效性，以解决大型复杂模型在训练和部署中面临的挑战。研究的核心问题是，随着VL...",
      "extracted_excerpt": "本文探讨了使用稀疏门控专家混合（MoE）技术来扩展视觉-语言模型（VLM）的有效性，以解决大型复杂模型在训练和部署中面临的挑战。研究的核心问题是，随着VL...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2024/04/01/shortcut-connected-expert-parallelism-for-accelerating-mixture-of-experts.html",
      "title": "Shortcut-connected Expert Parallelism for Accelerating Mixture of Experts",
      "raw_excerpt": "核心问题: 专家并行（Expert Parallelism）是分配稀疏门控混合专家（MoE）模型计算负载的关键策略，但其固有的All-to-All通信会产...",
      "extracted_excerpt": "核心问题: 专家并行（Expert Parallelism）是分配稀疏门控混合专家（MoE）模型计算负载的关键策略，但其固有的All-to-All通信会产...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2024/10/01/moe-accelerating-mixture-of-experts-methods-with-zero-computation-experts.html",
      "title": "MOE++: ACCELERATING MIXTURE-OF-EXPERTS METHODS WITH ZERO-COMPUTATION EXPERTS",
      "raw_excerpt": "本文旨在同时提升混合专家（Mixture-of-Experts, MoE）方法的有效性和效率。为实现此目标，本文提出了MoE++，一个集成了前馈网络（FF...",
      "extracted_excerpt": "本文旨在同时提升混合专家（Mixture-of-Experts, MoE）方法的有效性和效率。为实现此目标，本文提出了MoE++，一个集成了前馈网络（FF...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2022/01/01/deepspeed-moe-advancing-mixture-of-experts-inference-and-training-to-power-next-generation-ai-scale.html",
      "title": "DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale",
      "raw_excerpt": "本文旨在解决混合专家（Mixture-of-Experts, MoE）模型在实际应用中面临的挑战，使其成为训练和部署下一代大规模AI模型的可行且高效的方案。",
      "extracted_excerpt": "本文旨在解决混合专家（Mixture-of-Experts, MoE）模型在实际应用中面临的挑战，使其成为训练和部署下一代大规模AI模型的可行且高效的方案。",
      "excerpt_length": 78,
      "extracted_length": 78,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2024/10/01/eps-moe-expert-pipeline-scheduler-for-cost-efficient-moe-inference.html",
      "title": "EPS-MoE: Expert Pipeline Scheduler for Cost-Efficient MoE Inference",
      "raw_excerpt": "核心问题与研究目标\n大型语言模型（LLM）领域中，专家混合（MoE）模型已成为一种重要架构，它在模型性能和计算效率之间提供了更好的平衡。然而，MoE模型中...",
      "extracted_excerpt": "核心问题与研究目标\n大型语言模型（LLM）领域中，专家混合（MoE）模型已成为一种重要架构，它在模型性能和计算效率之间提供了更好的平衡。然而，MoE模型中...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2025/01/01/popfetcher-towards-accelerated-mixture-of-experts-training-via-popularity-based-expert-wise-prefetch.html",
      "title": "PopFetcher: Towards Accelerated Mixture-of-Experts Training Via Popularity Based Expert-Wise Prefetch",
      "raw_excerpt": "本文旨在解决混合专家（Mixture-of-Experts, MoE）模型训练中的通信和计算瓶颈。随着模型参数规模扩展至万亿级别，MoE架构虽然通过稀疏激...",
      "extracted_excerpt": "本文旨在解决混合专家（Mixture-of-Experts, MoE）模型训练中的通信和计算瓶颈。随着模型参数规模扩展至万亿级别，MoE架构虽然通过稀疏激...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2024/08/01/auxiliary-loss-free-load-balancing-strategy-for-mixture-of-experts.html",
      "title": "AUXILIARY-LOSS-FREE LOAD BALANCING STRATEGY FOR MIXTURE-OF-EXPERTS",
      "raw_excerpt": "本文旨在解决训练混合专家模型（Mixture-of-Experts, MoE）时面临的核心挑战：专家负载不均衡问题。该问题可能导致路由坍塌（部分专家训练不...",
      "extracted_excerpt": "本文旨在解决训练混合专家模型（Mixture-of-Experts, MoE）时面临的核心挑战：专家负载不均衡问题。该问题可能导致路由坍塌（部分专家训练不...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2021/12/01/glam-efficient-scaling-of-language-models-with-mixture-of-experts.html",
      "title": "GLaM: Efficient Scaling of Language Models with Mixture-of-Experts",
      "raw_excerpt": "本文提出并开发了一个名为GLaM（Generalist Language Model）的语言模型家族，旨在解决扩大语言模型（如GPT-3）带来的高昂计算资...",
      "extracted_excerpt": "本文提出并开发了一个名为GLaM（Generalist Language Model）的语言模型家族，旨在解决扩大语言模型（如GPT-3）带来的高昂计算资...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/moe/2022/06/01/tutel-adaptive-mixture-of-experts-at-scale.html",
      "title": "TUTEL: ADAPTIVE MIXTURE-OF-EXPERTS AT SCALE",
      "raw_excerpt": "本文针对稀疏门控混合专家（MoE）模型在扩展至万亿参数规模时面临的系统挑战，提出了一种名为 TUTEL 的高度可扩展的堆栈设计与实现。",
      "extracted_excerpt": "本文针对稀疏门控混合专家（MoE）模型在扩展至万亿参数规模时面临的系统挑战，提出了一种名为 TUTEL 的高度可扩展的堆栈设计与实现。",
      "excerpt_length": 67,
      "extracted_length": 67,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2025/04/01/tilelang-a-composable-tiled-programming-model-for-ai-systems.html",
      "title": "TileLang: A Composable Tiled Programming Model for AI Systems",
      "raw_excerpt": "现代AI工作负载严重依赖优化的计算核函数（kernel）进行训练和推理。尽管这些核函数遵循明确的数据流模式（如在DRAM和SRAM之间移动数据块并进行计算...",
      "extracted_excerpt": "现代AI工作负载严重依赖优化的计算核函数（kernel）进行训练和推理。尽管这些核函数遵循明确的数据流模式（如在DRAM和SRAM之间移动数据块并进行计算...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2025/04/01/tilelink-generating-efficient-compute-communication-overlapping-kernels-using-tile-centric-primitives.html",
      "title": "TileLink: Generating Efficient Compute-Communication Overlapping Kernels using Tile-Centric Primitives",
      "raw_excerpt": "本文旨在解决大型深度学习模型分布式训练中，通信开销（占总执行开销的10%-50%）限制计算效率的核心问题。提升性能最有效的方法是计算与通信的重叠（Over...",
      "extracted_excerpt": "本文旨在解决大型深度学习模型分布式训练中，通信开销（占总执行开销的10%-50%）限制计算效率的核心问题。提升性能最有效的方法是计算与通信的重叠（Over...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2018/10/01/tvm-an-automated-end-to-end-optimizing-compiler-for-deep-learning.html",
      "title": "TVM: An Automated End-to-End Optimizing Compiler for Deep Learning",
      "raw_excerpt": "核心问题与研究目标：随着机器学习应用需求的增长，需要将其部署到多样化的硬件设备上，从云服务器到自动驾驶汽车和嵌入式设备。然而，当前的深度学习（DL）框架（...",
      "extracted_excerpt": "核心问题与研究目标：随着机器学习应用需求的增长，需要将其部署到多样化的硬件设备上，从云服务器到自动驾驶汽车和嵌入式设备。然而，当前的深度学习（DL）框架（...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2024/10/01/flux-fast-software-based-communication-overlap-on-gpus-through-kernel-fusion.html",
      "title": "FLUX: FAST SOFTWARE-BASED COMMUNICATION OVERLAP ON GPUS THROUGH KERNEL FUSION",
      "raw_excerpt": "本文旨在解决大规模深度学习模型中张量并行（Tensor Parallelism）引入的通信开销问题。张量并行虽然能克服单个处理器内存容量限制并加速计算，但...",
      "extracted_excerpt": "本文旨在解决大规模深度学习模型中张量并行（Tensor Parallelism）引入的通信开销问题。张量并行虽然能克服单个处理器内存容量限制并加速计算，但...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2024/12/01/flex-attention-a-programming-model-for-generating-optimized-attention-kernels.html",
      "title": "FLEX ATTENTION: A PROGRAMMING MODEL FOR GENERATING OPTIMIZED ATTENTION KERNELS",
      "raw_excerpt": "本文旨在解决当前深度学习中注意力机制优化所面临的一个核心问题：高性能的注意力核（如FlashAttention）虽然效率极高，但其单一固化的设计（mono...",
      "extracted_excerpt": "本文旨在解决当前深度学习中注意力机制优化所面临的一个核心问题：高性能的注意力核（如FlashAttention）虽然效率极高，但其单一固化的设计（mono...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2025/06/01/mirage-a-multi-level-superoptimizer-for-tensor-programs.html",
      "title": "Mirage: A Multi-Level Superoptimizer for Tensor Programs",
      "raw_excerpt": "本文介绍了一种名为Mirage的多级张量程序超级优化器。当前用于深度神经网络（DNN）的框架，如PyTorch和TensorFlow，通常依赖手动设计的规...",
      "extracted_excerpt": "本文介绍了一种名为Mirage的多级张量程序超级优化器。当前用于深度神经网络（DNN）的框架，如PyTorch和TensorFlow，通常依赖手动设计的规...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/llm/engineering/compiler/2025/05/01/flashtensor-optimizing-tensor-programs-by-leveraging-fine-grained-tensor-property.html",
      "title": "FlashTensor: Optimizing Tensor Programs by Leveraging Fine-grained Tensor Property",
      "raw_excerpt": "核心问题: 深度神经网络（DNNs）在处理长上下文任务时会产生巨大的中间张量，导致严重的内存开销。现有的DNN优化方法由于对张量属性的感知不足，无法有效进...",
      "extracted_excerpt": "核心问题: 深度神经网络（DNNs）在处理长上下文任务时会产生巨大的中间张量，导致严重的内存开销。现有的DNN优化方法由于对张量属性的感知不足，无法有效进...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2025/09/01/categorical-foundations-for-cute-layouts.html",
      "title": "Categorical Foundations for CuTe Layouts",
      "raw_excerpt": "在现代计算，尤其是在 GPU 编程中，性能严重依赖于多维数据在内存中的存储和访问方式。尽管图像、视频和机器学习中的张量本质上是多维的，但计算机内存根本上是...",
      "extracted_excerpt": "在现代计算，尤其是在 GPU 编程中，性能严重依赖于多维数据在内存中的存储和访问方式。尽管图像、视频和机器学习中的张量本质上是多维的，但计算机内存根本上是...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2024/07/01/nvidia-blackwell-architecture-technical-brief.html",
      "title": "NVIDIA Blackwell Architecture Technical Brief",
      "raw_excerpt": "核心问题与研究目标\n随着人工智能（AI）和大型语言模型（LLM）的快速发展，实现实时性能和可扩展性变得至关重要。各行各业对生成式AI解决方案的需求激增，推...",
      "extracted_excerpt": "核心问题与研究目标\n随着人工智能（AI）和大型语言模型（LLM）的快速发展，实现实时性能和可扩展性变得至关重要。各行各业对生成式AI解决方案的需求激增，推...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2024/06/01/mind-the-gap-attainable-data-movement-and-operational-intensity-bounds-for-tensor-algorithms.html",
      "title": "Mind the Gap: Attainable Data Movement and Operational Intensity Bounds for Tensor Algorithms",
      "raw_excerpt": "核心问题：随着技术发展，数据移动对应用的性能和能效影响日益显著，尤其是在张量代数算法中。然而，当前用于早期架构设计空间探索（DSE）的数据移动评估方法存在...",
      "extracted_excerpt": "核心问题：随着技术发展，数据移动对应用的性能和能效影响日益显著，尤其是在张量代数算法中。然而，当前用于早期架构设计空间探索（DSE）的数据移动评估方法存在...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2018/04/01/gpudirect-async-exploring-gpu-synchronous-communication-techniques-for-infiniband-clusters.html",
      "title": "GPUDirect Async: Exploring GPU synchronous communication techniques for InfiniBand clusters",
      "raw_excerpt": "本文探讨了GPUDirect Async技术的动机和构建块，该技术是CUDA 8.0引入的新功能，允许GPU直接与第三方设备同步，例如直接触发和轮询排队到...",
      "extracted_excerpt": "本文探讨了GPUDirect Async技术的动机和构建块，该技术是CUDA 8.0引入的新功能，允许GPU直接与第三方设备同步，例如直接触发和轮询排队到...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2022/01/01/nvidia-h100-tensor-core-gpu-architecture.html",
      "title": "NVIDIA H100 Tensor Core GPU Architecture",
      "raw_excerpt": "本文档介绍了NVIDIA H100 Tensor Core GPU，这是NVIDIA基于Hopper GPU架构的新一代数据中心GPU。",
      "extracted_excerpt": "本文档介绍了NVIDIA H100 Tensor Core GPU，这是NVIDIA基于Hopper GPU架构的新一代数据中心GPU。",
      "excerpt_length": 68,
      "extracted_length": 68,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2024/03/01/wasp-exploiting-gpu-pipeline-parallelism-with-hardware-accelerated-automatic-warp-specialization.html",
      "title": "WASP: Exploiting GPU Pipeline Parallelism with Hardware-Accelerated Automatic Warp Specialization",
      "raw_excerpt": "本文介绍了一种名为WASP的架构和编译器，旨在解决现代GPU中warp专业化（warp specialization）技术的三大局限性，以提升对内存延迟和...",
      "extracted_excerpt": "本文介绍了一种名为WASP的架构和编译器，旨在解决现代GPU中warp专业化（warp specialization）技术的三大局限性，以提升对内存延迟和...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2023/01/01/a-135-gbpsgbit-066-pjbit-stacked-embedded-dram-with-multilayer-arrays-by-fine-pitch-hybrid-bonding-and-mini-tsv.html",
      "title": "A 135 GBps/Gbit 0.66 pJ/bit Stacked Embedded DRAM with Multilayer Arrays by Fine Pitch Hybrid Bonding and Mini-TSV",
      "raw_excerpt": "人工智能（AI）和图形处理单元（GPU）技术的快速发展增加了对并行处理的需求，从而提高了对内存带宽的需求。图形双数据率内存（如GDDR6【1, K. D....",
      "extracted_excerpt": "人工智能（AI）和图形处理单元（GPU）技术的快速发展增加了对并行处理的需求，从而提高了对内存带宽的需求。图形双数据率内存（如GDDR6【1, K. D....",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2023/05/01/hardware-compute-partitioning-on-nvidia-gpus.html",
      "title": "Hardware Compute Partitioning on NVIDIA GPUs\\*",
      "raw_excerpt": "本文旨在解决当前 GPU 管理方法中存在的一个核心问题：GPU 通常被视为一个单一的、不可分割的设备，并强制执行互斥访问控制，这类似于在多核 CPU 上一...",
      "extracted_excerpt": "本文旨在解决当前 GPU 管理方法中存在的一个核心问题：GPU 通常被视为一个单一的、不可分割的设备，并强制执行互斥访问控制，这类似于在多核 CPU 上一...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2016/04/01/optimizing-performance-of-recurrent-neural-networks-on-gpus.html",
      "title": "Optimizing Performance of Recurrent Neural Networks on GPUs",
      "raw_excerpt": "随着循环神经网络（RNN）变得越来越大和越来越深，单个网络的训练时间上升到数周甚至数月，因此提高这些网络的性能和可扩展性具有重要意义。尽管GPU已成为训练...",
      "extracted_excerpt": "随着循环神经网络（RNN）变得越来越大和越来越深，单个网络的训练时间上升到数周甚至数月，因此提高这些网络的性能和可扩展性具有重要意义。尽管GPU已成为训练...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2023/01/01/stream-k-work-centric-parallel-decomposition-for-dense-matrix-matrix-multiplication-on-the-gpu.html",
      "title": "Stream-K: Work-centric Parallel Decomposition for Dense Matrix-Matrix Multiplication on the GPU",
      "raw_excerpt": "本文介绍了一种名为 Stream-K 的新颖并行分解方法，该方法以工作为中心，用于稠密线性代数中的矩阵乘法（GEMM）及相关计算。与当前主要基于分块（ti...",
      "extracted_excerpt": "本文介绍了一种名为 Stream-K 的新颖并行分解方法，该方法以工作为中心，用于稠密线性代数中的矩阵乘法（GEMM）及相关计算。与当前主要基于分块（ti...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2009/04/01/roofline-an-insightful-visual-performance-model-for-multicore-architectures.html",
      "title": "Roofline: An Insightful Visual Performance Model for Multicore Architectures",
      "raw_excerpt": "核心问题: 计算机体系结构从单核向多核的转变带来了设计上的多样性，例如，众多的简单处理器与少数复杂处理器、多线程技术的依赖程度、以及使用显式寻址的本地存储...",
      "extracted_excerpt": "核心问题: 计算机体系结构从单核向多核的转变带来了设计上的多样性，例如，众多的简单处理器与少数复杂处理器、多线程技术的依赖程度、以及使用显式寻址的本地存储...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2017/05/01/offloading-communication-control-logic-in-gpu-accelerated-applications.html",
      "title": "Offloading communication control logic in GPU accelerated applications",
      "raw_excerpt": "本文从引言中提取的核心问题是：在多GPU加速的应用中，传统的GPUDirect技术主要关注数据移动的效率，但控制路径仍依赖CPU，导致CPU在计算和通信阶...",
      "extracted_excerpt": "本文从引言中提取的核心问题是：在多GPU加速的应用中，传统的GPUDirect技术主要关注数据移动的效率，但控制路径仍依赖CPU，导致CPU在计算和通信阶...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2010/01/01/demystifying-gpu-microarchitecture-through-microbenchmarking.html",
      "title": "Demystifying GPU Microarchitecture through Microbenchmarking",
      "raw_excerpt": "本文旨在通过开发一套微基准测试程序，来测量和揭示Nvidia GT200 (GTX280) GPU中对CUDA可见的架构特性。由于制造商提供的文档信息有限...",
      "extracted_excerpt": "本文旨在通过开发一套微基准测试程序，来测量和揭示Nvidia GT200 (GTX280) GPU中对CUDA可见的架构特性。由于制造商提供的文档信息有限...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2025/03/01/neutrino-fine-grained-gpu-kernel-profiling-via-programmable-probing.html",
      "title": "Neutrino: Fine-grained GPU Kernel Profiling via Programmable Probing",
      "raw_excerpt": "本论文介绍了NEUTRINO，这是一个可编程的GPU内核剖析接口，通过汇编层探测实现指令级细粒度剖析、在时间和值域上的剖析多功能性，以及硬件独立性。为了更...",
      "extracted_excerpt": "本论文介绍了NEUTRINO，这是一个可编程的GPU内核剖析接口，通过汇编层探测实现指令级细粒度剖析、在时间和值域上的剖析多功能性，以及硬件独立性。为了更...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2017/04/01/locality-aware-cta-clustering-for-modern-gpus.html",
      "title": "Locality-Aware CTA Clustering for Modern GPUs",
      "raw_excerpt": "本文探讨了现代GPU中一种长期被忽视但具有性能提升潜力的数据局部性——CTA间局部性（inter-CTA locality）。利用这种局部性面临三大挑战：...",
      "extracted_excerpt": "本文探讨了现代GPU中一种长期被忽视但具有性能提升潜力的数据局部性——CTA间局部性（inter-CTA locality）。利用这种局部性面临三大挑战：...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2025/07/01/dissecting-the-nvidia-blackwell-architecture-with-microbenchmarks.html",
      "title": "Dissecting the NVIDIA Blackwell Architecture with Microbenchmarks",
      "raw_excerpt": "本文旨在通过微基准测试对NVIDIA Blackwell GPU架构进行微架构层面的分析，并与前代Hopper架构进行比较。研究的核心问题在于，面对层出不...",
      "extracted_excerpt": "本文旨在通过微基准测试对NVIDIA Blackwell GPU架构进行微架构层面的分析，并与前代Hopper架构进行比较。研究的核心问题在于，面对层出不...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2025/01/01/nvidia-rtx-blackwell-gpu-architecture.html",
      "title": "NVIDIA RTX BLACKWELL GPU ARCHITECTURE",
      "raw_excerpt": "NVIDIA RTX Blackwell 架构建立在先前 NVIDIA GPU 引入的基础 AI 技术之上，旨在为下一代 AI 驱动的游戏和专业应用提供支...",
      "extracted_excerpt": "NVIDIA RTX Blackwell 架构建立在先前 NVIDIA GPU 引入的基础 AI 技术之上，旨在为下一代 AI 驱动的游戏和专业应用提供支...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2018/03/01/improving-real-time-performance-with-cuda-persistent-threads-cuper-on-the-jetson-tx2.html",
      "title": "Improving Real-Time Performance with CUDA Persistent Threads (CuPer) on the Jetson TX2",
      "raw_excerpt": "本文旨在解决将图形处理单元（GPU）用于具有亚毫秒级帧时长约束的硬实时应用时遇到的性能确定性差的问题。传统的CUDA启动/同步编程模型会引入显著的抖动，不...",
      "extracted_excerpt": "本文旨在解决将图形处理单元（GPU）用于具有亚毫秒级帧时长约束的硬实时应用时遇到的性能确定性差的问题。传统的CUDA启动/同步编程模型会引入显著的抖动，不...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2025/06/01/serving-large-language-models-on-huawei-cloudmatrix384.html",
      "title": "Serving Large Language Models on Huawei CloudMatrix384",
      "raw_excerpt": "本文针对大型语言模型（LLM）因参数规模增长、专家混合（MoE）架构采用和上下文长度扩展而对AI基础设施带来的前所未有的挑战，介绍了华为CloudMatr...",
      "extracted_excerpt": "本文针对大型语言模型（LLM）因参数规模增长、专家混合（MoE）架构采用和上下文长度扩展而对AI基础设施带来的前所未有的挑战，介绍了华为CloudMatr...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/gpu/2020/02/01/gpu-initiated-openshmem-correct-and-eicient-intra-kernel-networking-for-dgpus.html",
      "title": "GPU Initiated OpenSHMEM: Correct and Eicient Intra-Kernel Networking for dGPUs",
      "raw_excerpt": "本文旨在解决当前GPU网络通信中存在的性能瓶颈和编程复杂性问题。",
      "extracted_excerpt": "本文旨在解决当前GPU网络通信中存在的性能瓶颈和编程复杂性问题。",
      "excerpt_length": 32,
      "extracted_length": 32,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2019/10/01/netdimm-low-latency-near-memory-network-interface-architecture.html",
      "title": "NetDIMM: Low-Latency Near-Memory Network Interface Architecture",
      "raw_excerpt": "本文旨在解决数据中心网络中两个核心瓶颈：PCIe互连的高延迟开销和网络软件栈中的数据移动开销。传统的网络设计主要关注带宽，但现代数据中心应用（如内存缓存、...",
      "extracted_excerpt": "本文旨在解决数据中心网络中两个核心瓶颈：PCIe互连的高延迟开销和网络软件栈中的数据移动开销。传统的网络设计主要关注带宽，但现代数据中心应用（如内存缓存、...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2025/01/01/autoccl-automated-collective-communication-tuning-for-accelerating-distributed-and-parallel-dnn-training.html",
      "title": "AutoCCL: Automated Collective Communication Tuning for Accelerating Distributed and Parallel DNN Training",
      "raw_excerpt": "本文针对分布式并行深度神经网络（DNN）训练中通信瓶颈问题，提出了一种名为AutoCCL的自动化集合通信调优方法。现有网络优化研究通常假设底层通信库（如N...",
      "extracted_excerpt": "本文针对分布式并行深度神经网络（DNN）训练中通信瓶颈问题，提出了一种名为AutoCCL的自动化集合通信调优方法。现有网络优化研究通常假设底层通信库（如N...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2023/07/01/rail-only-a-low-cost-high-performance-network-for-training-llms-with-trillion-parameters.html",
      "title": "Rail-only: A Low-Cost High-Performance Network for Training LLMs with Trillion Parameters",
      "raw_excerpt": "本文针对超大规模大型语言模型（LLM）的训练，提出了一种低成本的网络架构。随着LLM模型规模和计算需求的增长速度超过硬件发展，超大规模GPU数据中心成为必...",
      "extracted_excerpt": "本文针对超大规模大型语言模型（LLM）的训练，提出了一种低成本的网络架构。随着LLM模型规模和计算需求的增长速度超过硬件发展，超大规模GPU数据中心成为必...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2023/07/01/overview-of-and-motivation-for-the-forthcoming-ultra-ethernet-consortium-specification.html",
      "title": "Overview of and Motivation for the Forthcoming Ultra Ethernet Consortium Specification",
      "raw_excerpt": "本文作为一份白皮书，阐述了为满足现代人工智能（AI）和高性能计算（HPC）工作负载需求而成立的 Ultra Ethernet 联盟（UEC）及其即将推出的...",
      "extracted_excerpt": "本文作为一份白皮书，阐述了为满足现代人工智能（AI）和高性能计算（HPC）工作负载需求而成立的 Ultra Ethernet 联盟（UEC）及其即将推出的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2022/11/01/improving-network-performance-of-hpc-systems-using-nvidia-magnum-io-nvshmem-and-gpudirect-async.html",
      "title": "Improving Network Performance of HPC Systems Using NVIDIA Magnum IO NVSHMEM and GPUDirect Async",
      "raw_excerpt": "本文旨在解决当前大规模高性能计算（HPC）系统中存在的细粒度通信效率低下的核心问题。随着HPC系统扩展到数万个GPU，特别是在强缩放（strong sca...",
      "extracted_excerpt": "本文旨在解决当前大规模高性能计算（HPC）系统中存在的细粒度通信效率低下的核心问题。随着HPC系统扩展到数万个GPU，特别是在强缩放（strong sca...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2022/02/01/doubling-all2all-performance-with-nvidia-collective-communication-library-212.html",
      "title": "Doubling all2all Performance with NVIDIA Collective Communication Library 2.12",
      "raw_excerpt": "本文主要介绍了NVIDIA Collective Communication Library (NCCL) 2.12版本中的一项关键性能优化。",
      "extracted_excerpt": "本文主要介绍了NVIDIA Collective Communication Library (NCCL) 2.12版本中的一项关键性能优化。",
      "excerpt_length": 71,
      "extracted_length": 71,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2014/11/01/introducing-data-center-fabric-the-next-generation-facebook-data-center-network.html",
      "title": "Introducing data center fabric, the next-generation Facebook data center network",
      "raw_excerpt": "本文介绍了Facebook为其新一代数据中心（从Altoona设施开始）设计并成功部署的全新网络架构——数据中心Fabric。",
      "extracted_excerpt": "本文介绍了Facebook为其新一代数据中心（从Altoona设施开始）设计并成功部署的全新网络架构——数据中心Fabric。",
      "excerpt_length": 63,
      "extracted_length": 63,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2015/07/01/ucx-an-open-source-framework-for-hpc-network-apis-and-beyond.html",
      "title": "UCX: An Open Source Framework for HPC Network APIs and Beyond",
      "raw_excerpt": "本文针对高性能计算（HPC）领域中网络编程模型的可移植性与性能挑战，提出了一个名为统一通信X（Unified Communication X, UCX）的...",
      "extracted_excerpt": "本文针对高性能计算（HPC）领域中网络编程模型的可移植性与性能挑战，提出了一个名为统一通信X（Unified Communication X, UCX）的...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2016/06/01/design-guidelines-for-high-performance-rdma-systems.html",
      "title": "Design Guidelines for High Performance RDMA Systems",
      "raw_excerpt": "本文旨在为系统设计者提供一套设计指南，以帮助他们在RDMA（远程直接内存访问）的设计空间中做出最佳选择。尽管现代RDMA硬件性能强大，但如何选择和使用RD...",
      "extracted_excerpt": "本文旨在为系统设计者提供一套设计指南，以帮助他们在RDMA（远程直接内存访问）的设计空间中做出最佳选择。尽管现代RDMA硬件性能强大，但如何选择和使用RD...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2015/08/01/jupiter-rising-a-decade-of-clos-topologies-and-centralized-control-in-googles-datacenter-network.html",
      "title": "Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network",
      "raw_excerpt": "本文回顾了过去十年间，Google为克服数据中心网络在成本、运营复杂性和规模限制方面的问题而采取的策略。论文详细介绍了五代数据中心网络的发展历程，这些网络...",
      "extracted_excerpt": "本文回顾了过去十年间，Google为克服数据中心网络在成本、运营复杂性和规模限制方面的问题而采取的策略。论文详细介绍了五代数据中心网络的发展历程，这些网络...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2025/07/01/demystifying-nccl-an-in-depth-analysis-of-gpu-communication-protocols-and-algorithms.html",
      "title": "Demystifying NCCL: An In-depth Analysis of GPU Communication Protocols and\n      Algorithms",
      "raw_excerpt": "本文针对NVIDIA Collective Communication Library\n        (NCCL)的内部设计进行全面分析，核心问题在于N...",
      "extracted_excerpt": "本文针对NVIDIA Collective Communication Library\n        (NCCL)的内部设计进行全面分析，核心问题在于N...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2025/04/01/introducing-ualink-200g-10-specification.html",
      "title": "Introducing UALink 200G 1.0 Specification",
      "raw_excerpt": "本文介绍了Ultra Accelerator Link™ (UALink™) 200G 1.0 规范，这是一个为下一代人工智能工作负载设计的全新行业标准，...",
      "extracted_excerpt": "本文介绍了Ultra Accelerator Link™ (UALink™) 200G 1.0 规范，这是一个为下一代人工智能工作负载设计的全新行业标准，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2025/03/01/ub-mesh-a-hierarchically-localized-nd-fullmesh-datacenter-network-architecture.html",
      "title": "UB-Mesh: a Hierarchically Localized nD-FullMesh Datacenter Network Architecture",
      "raw_excerpt": "随着大型语言模型（LLM）的持续扩展，其对计算能力和带宽的需求也随之升级。本文针对下一代AI数据中心面临的大规模（R1）、高带宽（R2）、成本效益（R3）...",
      "extracted_excerpt": "随着大型语言模型（LLM）的持续扩展，其对计算能力和带宽的需求也随之升级。本文针对下一代AI数据中心面临的大规模（R1）、高带宽（R2）、成本效益（R3）...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2020/08/01/an-in-depth-analysis-of-the-slingshot-interconnect.html",
      "title": "An In-Depth Analysis of the Slingshot Interconnect",
      "raw_excerpt": "本文深入分析了为大规模计算系统设计的 Slingshot 互连网络。随着系统规模的扩大，互连网络对应用性能的影响日益关键，特别是在高性能计算（HPC）领域...",
      "extracted_excerpt": "本文深入分析了为大规模计算系统设计的 Slingshot 互连网络。随着系统规模的扩大，互连网络对应用性能的影响日益关键，特别是在高性能计算（HPC）领域...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2022/07/01/reexamining-direct-cache-access-to-optimize-io-intensive-applications-for-multi-hundred-gigabit-networks.html",
      "title": "Reexamining Direct Cache Access to Optimize I/O Intensive Applications for Multi-hundred-gigabit Networks",
      "raw_excerpt": "本文旨在研究缓存管理对I/O密集型应用性能的影响，特别关注数据包处理中的一个瓶颈：直接缓存访问（Direct Cache Access, DCA）。随着网...",
      "extracted_excerpt": "本文旨在研究缓存管理对I/O密集型应用性能的影响，特别关注数据包处理中的一个瓶颈：直接缓存访问（Direct Cache Access, DCA）。随着网...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2025/07/01/scale-up-ethernet-framework-scale-up-ethernet-framework-specification.html",
      "title": "Scale-Up Ethernet Framework Scale-Up Ethernet Framework Specification",
      "raw_excerpt": "本文档介绍了可扩展以太网框架（Scale-Up Ethernet, SUE），旨在为基于以太网的XPU（如图形处理器GPU或各种定制机器学习加速器）横向扩...",
      "extracted_excerpt": "本文档介绍了可扩展以太网框架（Scale-Up Ethernet, SUE），旨在为基于以太网的XPU（如图形处理器GPU或各种定制机器学习加速器）横向扩...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/networking/2024/04/01/scaling-up-memory-disaggregated-applications-with-smart.html",
      "title": "Scaling Up Memory Disaggregated Applications with Smart",
      "raw_excerpt": "本文旨在解决IOPS密集型的内存分离应用在现代多核服务器上扩展性不佳的问题。研究发现，当核心数超过32个时，这类应用的吞吐量通常会达到瓶颈并开始下降，无法...",
      "extracted_excerpt": "本文旨在解决IOPS密集型的内存分离应用在现代多核服务器上扩展性不佳的问题。研究发现，当核心数超过32个时，这类应用的吞吐量通常会达到瓶颈并开始下降，无法...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/cpu/2022/05/01/everything-you-need-to-know-about-the-cpu-power-management.html",
      "title": "Everything You Need to Know About the CPU Power Management",
      "raw_excerpt": "本文的核心目标是详细阐述CPU的多种低功耗模式，即C-states，以帮助读者理解其工作原理和节能效果。",
      "extracted_excerpt": "本文的核心目标是详细阐述CPU的多种低功耗模式，即C-states，以帮助读者理解其工作原理和节能效果。",
      "excerpt_length": 52,
      "extracted_length": 52,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/cpu/2022/05/01/understanding-bios-configuration-for-performance-tuning.html",
      "title": "Understanding BIOS Configuration for Performance Tuning",
      "raw_excerpt": "本文档旨在指导用户如何通过调整BIOS（基本输入/输出系统）设置来优化服务器性能。",
      "extracted_excerpt": "本文档旨在指导用户如何通过调整BIOS（基本输入/输出系统）设置来优化服务器性能。",
      "excerpt_length": 41,
      "extracted_length": 41,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/compiler/2025/01/01/pipethreader-software-defined-pipelining-for-efficient-dnn-execution.html",
      "title": "PipeThreader: Software-Defined Pipelining for Efficient DNN Execution",
      "raw_excerpt": "本文提出PipeThreader，一个新的DNN编译器，旨在有效利用现代GPU中的异构专用硬件单元，如TensorCores和Tensor Memory ...",
      "extracted_excerpt": "本文提出PipeThreader，一个新的DNN编译器，旨在有效利用现代GPU中的异构专用硬件单元，如TensorCores和Tensor Memory ...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/compiler/2022/07/01/alpa-automating-inter-and-intra-operator-parallelism-for-distributed-deep-learning.html",
      "title": "Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning",
      "raw_excerpt": "核心问题: 训练如GPT-3等拥有数千亿参数的超大规模深度学习模型，需要大量的工程努力，这些努力高度依赖于特定的模型定义和集群环境。例如，为Transfo...",
      "extracted_excerpt": "核心问题: 训练如GPT-3等拥有数千亿参数的超大规模深度学习模型，需要大量的工程努力，这些努力高度依赖于特定的模型定义和集群环境。例如，为Transfo...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/compiler/2025/05/01/kperfir-towards-an-open-and-compiler-centric-ecosystem-for-gpu-kernel-performance-tooling-on-modern-ai-workloads.html",
      "title": "KPerfIR: Towards an Open and Compiler-centric Ecosystem for GPU Kernel Performance Tooling on Modern AI Workloads",
      "raw_excerpt": "本工作提出KPerfIR，一种新型的多级编译器中心基础设施，用于支持在现代GPU上针对现代人工智能（AI）工作负载的自定义、可扩展和可移植的性能分析工具开...",
      "extracted_excerpt": "本工作提出KPerfIR，一种新型的多级编译器中心基础设施，用于支持在现代GPU上针对现代人工智能（AI）工作负载的自定义、可扩展和可移植的性能分析工具开...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/framework/2022/01/01/campo-cost-aware-performance-optimization-for-mixed-precision-neural-network-training.html",
      "title": "Campo: Cost-Aware Performance Optimization for Mixed-Precision Neural Network Training",
      "raw_excerpt": "本文的核心研究问题是，现有的混合精度训练策略（如TensorFlow和PyTorch中的策略）贪婪地对性能关键型操作应用FP16精度，但没有量化和考虑数据...",
      "extracted_excerpt": "本文的核心研究问题是，现有的混合精度训练策略（如TensorFlow和PyTorch中的策略）贪婪地对性能关键型操作应用FP16精度，但没有量化和考虑数据...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/mlsys/system/2025/01/01/principles-and-methodologies-for-serial-performance-optimization.html",
      "title": "Principles and Methodologies for Serial Performance Optimization",
      "raw_excerpt": "本文针对计算机系统中顺序执行性能优化的长期挑战，提出了一种系统化的方法。核心问题是顺序任务的优化对整体性能至关重要，因为根据Amdahl定律，系统的最大加...",
      "extracted_excerpt": "本文针对计算机系统中顺序执行性能优化的长期挑战，提出了一种系统化的方法。核心问题是顺序任务的优化对整体性能至关重要，因为根据Amdahl定律，系统的最大加...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    },
    {
      "url": "/papers/diffusions/dllm/2025/07/01/fast-dllm-training-free-acceleration-of-diffusion-llm-by-enabling-kv-cache-and-parallel-decoding.html",
      "title": "Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding",
      "raw_excerpt": "本文旨在解决开源的基于扩散的大语言模型（Diffusion LLM）在实际推理速度上落后于自回归（AR）模型的核心问题。其性能瓶颈主要源于两个方面：首先，...",
      "extracted_excerpt": "本文旨在解决开源的基于扩散的大语言模型（Diffusion LLM）在实际推理速度上落后于自回归（AR）模型的核心问题。其性能瓶颈主要源于两个方面：首先，...",
      "excerpt_length": 80,
      "extracted_length": 80,
      "is_extracted": false
    }
  ]
}