Three paper indexes(LLM/Transactions/Distributed Systems) are pinned.
Other posts are sorted by date.
Three paper indexes(LLM/Transactions/Distributed Systems) are pinned.
Other posts are sorted by date.
My reading notes. 2026 0203-0209 Cascade Speculative Drafting for Even Faster LLM Inference CAS-Spec Cascade Adaptive Self-Speculative Decoding for On-the-Fly Lossless Inference Acceleration of LLMs Draft & Verify Lossless Large Language Model Acceleration via Self-Speculative Decoding Swift On-the-fly Self-speculative Decoding For LLM Inference Acceleration 3-Model Speculative Decoding Hierarchical Speculative Decoding with Dynamic Windows for Efficient Language Model Inference LayerSkip Enabling Early Exit Inference and Self-Speculative Decoding AIConfigurator Lightning-Fast Configuration Optimization for Multi-Framework LLM Serving Revati Transparent GPU-Free Time-Warp Emulation for LLM Serving 0127-0202 FlexPrefill A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference XAttention Block Sparse Attention with Antidiagonal Scoring SLED A Speculative LLM Decoding Framework for Efficient Edge Serving R-Stitch Dynamic Trajectory Stitching for Efficient Reasoning Estimating LLM Uncertainty with Evidence Entropy Adaptive Decoding Dynamic Model Switching for Efficient Inference Think Big, Generate Quick LLM-to-SLM for Fast Autoregressive Decoding 2025 Remaining Beyond the 80 20 Rule High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning KVCache Cache in the Wild Characterizing and Optimizing KVCache Cache at a Large Cloud Provider 1111-1117 LServe Efficient Long-sequence LLM Serving with Unified Sparse Attention QServe W4A8KV4 Quantization and System Co-design for Efficient LLM Serving Quest Query-Aware Sparsity for Efficient Long-Context LLM Inference Dynamic Sparse Attention on Mobile SoCs A dynamic parallel method for performance optimization on hybrid CPUs SmoothQuant Accurate and Efficient Post-Training Quantization for Large Language Models DuoAttention Efficient Long-Context LLM Inference with Retrieval and Streaming Heads Efficient Streaming Language Models with Attention Sinks KTransformers Unleashing the Full Potential of CPU GPU Hybrid Inference for MoE Models 1104-1110 EAGLE Speculative Sampling Requires Rethinking Feature Uncertainty 1028-1103 Aegaeon Effective GPU Pooling for Concurrent LLM Serving on the Market DistServe Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving Splitwise Efficient Generative LLM Inference Using Phase Splitting Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve 0826-0901 ELMS Elasticized Large Language Models On Mobile Devices Scaling Up On-Device LLMs via Active-Weight Swapping Between DRAM and Flash 0819-0825 STI Turbocharge NLP Inference at the Edge via Elastic Pipelining EdgeMoE Empowering Sparse Large Language Models on Mobile Devices LLM as a System Service on Mobile Devices SmallThinker A Family of Efficient Large Language Models Natively Trained for Local Deployment HeteroLLM Accelerating Large Language Model Inference on Mobile SoCs with Heterogeneous AI Accelerators A Survey of Resource-efficient LLM and Multimodal Foundation Models H2O Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models 0812-0818 KV-Runahead Scalable Causal LLM Inference by Parallel Key-Value Cache Generation Striped Attention Faster Ring Attention for Causal Transformers Ring Attention with Blockwise Transformers for Near-Infinite Context TPI-LLM Serving 70B-scale LLMs Efficiently on Low-resource Mobile Devices LLM.int8() 8-bit Matrix Multiplication for Transformers at Scale 0729-0804 Fast On-device LLM Inference with NPUs Deja Vu Contextual Sparsity for Efficient LLMs at Inference Time PowerInfer-2 Fast Large Language Model Inference on a Smartphone LLM in a flash Efficient Large Language Model Inference with Limited Memory PowerInfer Fast Large Language Model Serving with a Consumer-grade GPU 0722-0728 AWQ Activation-aware Weight Quantization for LLM Compression and Acceleration FlexGen High-Throughput Generative Inference of Large Language Models with a Single GPU LoRA Low-Rank Adaptation of Large Language Models SpecInfer Accelerating Large Language Model Serving with Tree-based Speculative Inference and Verification EdgeLLM Fast On-Device LLM Inference With Speculative Decoding Efficient Memory Management for Large Language Model Serving with PagedAttention 0715-0721 A Survey on Efficient Inference for Large Language Models -0714 Orca A Distributed Serving System for Transformer-Based Generative Models EdgeShard Efficient LLM Inference via Collaborative Edge Computing ServerlessLLM Locality-Enhanced Serverless Inference for Large Language Models Uncategorized WIP 🚧 ...
My reading notes. 2025 0715-0721 Concurrency Control as a Service Sonata Multi-Database Transactions Made Fast and Serializable Uncategorized WIP 🚧 towards-transaction-as-a-service grit taking-omid-to-the-clouds epoxy ad-hoc-transactions-in-web-applications omid-reloaded data-management-in-microservices scalable-distributed-transactions-across-heterogeneous-stores cobra
My reading notes. 2025 2023 && 2024 bigtable cap-twelve-years-later zab mapreduce chubby chain-replication time, clocks, and the ordering farm zookeeper
vLLM 中同时存在多条与 FX 和 torch.compile 相关的执行路径,它们虽然共享 PyTorch 的 graph infrastructure,但解决的问题并不相同。理解这些路径的边界,是区分“模型结构改写”“Inductor 编译优化”和“CUDA Graph replay”的前提。 进入正文前,可以先区分三类用法: HF Transformers modeling backend:直接使用 fx.Tracer 分析 Hugging Face forward,识别可融合结构并通过 AST 改写模型源码 vLLM 自定义 torch.compile 主路径:由 Dynamo 产生完整 FX GraphModule,再由 vLLM 控制 graph partition、shape specialization、custom passes、cache 和 CUDA Graph integration 普通 @torch.compile 小函数:FX 主要作为 Dynamo、AOTAutograd 和 Inductor 的内部 IR,vLLM 通常不直接处理这些 graph Piecewise CUDA Graph 不是第四条独立的 tracing 路径,而是建立在 vLLM Piecewise Compilation 之上的运行时优化:编译阶段先拆分并编译 graph regions,随后只对兼容区域执行 CUDA Graph capture/replay。 这篇文章的核心问题是:vLLM 如何围绕标准 torch.compile 增加一层 LLM-serving-aware 的 graph processing、compilation policy 和 runtime integration。 ...
核心结论 Piecewise CUDA Graph 的核心思想是:不要求整个 Transformer forward 都满足 CUDA Graph capture 条件,而是将 CUDA Graph 不兼容的算子作为边界,只 capture 其余 CUDA-Graph-safe 区域。 在 vLLM 中,最典型的边界是 Attention。因此,PIECEWISE 模式并不是“把一个 CUDA Graph 切开执行”,而是将模型划分为多个独立的 CUDA Graph,并在它们之间以 eager 模式执行 CUDA-Graph-unsafe 操作: CUDA Graph -> Attention (eager) -> CUDA Graph -> Attention (eager) -> CUDA Graph CUDA Graph 会提前 capture 一组 GPU operations,后续通过 graph replay 一次性提交,从而降低 CPU 逐个启动 kernel 的开销。Transformer 中的 RMSNorm、Linear/GEMM、activation、FFN 和 residual 等计算,在 capture size 确定时通常具有稳定的执行结构;Attention 则需要处理 KV Cache、动态 shape 和运行时 metadata,对 CUDA Graph 的兼容要求更高。 ...
ML Compiler Overview ML Compiler 位于深度学习框架和底层硬件之间,作用是把 TensorFlow、PyTorch、JAX 等框架描述的高层张量计算转换成能在 CPU/GPU/TPU 等硬件上高效执行的程序。 它的典型架构是 Frontend → IR → Optimization → Backend: Frontend 接收框架已经构造或捕获出来的计算图/算子程序(例如 tf.function 得到的 TF Graph、PyTorch Dynamo 得到的 FX Graph),转换成统一的中间表示 IR 中间层进行算子融合、常量折叠、布局变换、内存规划、并行化等与机器学习计算相关的优化 Backend 再根据目标硬件进行 lowering、调度和代码生成 因而 ML compiler 输入通常是带有 tensor shape、dtype、算子及数据依赖关系的计算图或 IR,输出则是面向特定设备的低层 IR、kernel 或可执行程序。 需要特别区分的是:graph capture 不一定属于 compiler 本身——例如 TensorFlow 由 tf.function tracing 捕获图,PyTorch 由 TorchDynamo 捕获图,然后 ML compiler 才接手,对这个图做优化并最终生成高效机器码。 IR Pipeline 下面这套 IR 划分描述的是一个典型 ML Compiler 从高层模型计算逐渐 lowering 到硬件可执行 kernel 的过程。不同编译器实际使用的 IR 名称和数据结构可能不同,但抽象层次通常可以归纳为: ...
在和 GGML 打交道时不知道已经写了几篇文档了,这篇一定是最后一个 这篇文档主要记录我用 GGML 实现一个简单的 LLM Inference Engine 时遇到的问题 C++ Initialization 在 C++ 中, class 中的成员在构造函数体开始之前就被默认构造了,要控制其行为,需要使用初始化列表 class Member { public: Member() { std::cout << "Member 默认构造\n"; } Member(int x) { std::cout << "Member 带参构造: " << x << "\n"; } }; class MyClass { Member m1; // 成员对象 Member m2; int value; public: // 情况1: 不使用初始化列表 MyClass() { std::cout << "构造函数体开始\n"; value = 10; // 这是赋值,不是初始化! } // 情况2: 使用初始化列表 MyClass(int v) : m1(1), m2(2), value(v) { std::cout << "构造函数体开始\n"; } }; 情况 1 输出 Member 默认构造 // m1 在构造函数体前被默认构造 Member 默认构造 // m2 在构造函数体前被默认构造 构造函数体开始 情况 2 输出 Member 带参构造: 1 // m1 在构造函数体前初始化 Member 带参构造: 2 // m2 在构造函数体前初始化 构造函数体开始 对于指针来说,初始化就是将其设置为 nullptr, 对于 STL 容器,就是初始化空容器 ...
向量内积 物理意义 内积本质上是一个向量在另一个向量方向上的投影长度,与基准向量长度的乘积。 如果提取特征或求取分量,令 $\mathbf{u}$ 为单位向量($\|\mathbf{u}\| = 1$),则 $\mathbf{v} \cdot \mathbf{u}$ 直接输出 $\mathbf{v}$ 在 $\mathbf{u}$ 方向上的标量投影。 工程应用: 在信号处理(如傅里叶变换)中,信号与正交基函数的内积,就是在提取该信号在特定频率上的能量分量。在经典力学中,功的计算 $W = \mathbf{F} \cdot \mathbf{d}$ 就是提取力在位移方向上的有效分量并相乘。 由于内积公式中包含 $\cos(\theta)$,它是衡量高维空间中两个向量方向“一致性”或“对齐程度”的线性算子。 当 $\mathbf{a} \cdot \mathbf{b} > 0$ 时,夹角为锐角,两者存在正相关性。 当 $\mathbf{a} \cdot \mathbf{b} = 0$ 时,$\cos(\theta) = 0$,两向量正交(垂直)。在工程上,这意味着两个系统、信号或特征完全独立,互不干涉(即协方差为零)。 当 $\mathbf{a} \cdot \mathbf{b} < 0$ 时,夹角为钝角,存在负相关性。 工程应用: 在机器学习和数据挖掘中,将向量归一化后求内积,即为余弦相似度(Cosine Similarity),常用于衡量文本词向量的语义相似性或推荐系统中用户偏好的匹配度。 Note 向量内积天然满足交换律,即 $\mathbf{a} \cdot \mathbf{b} = \mathbf{b} \cdot \mathbf{a}$,即 $\mathbf{a}$ 投影在 $\mathbf{b}$ 上和 $\mathbf{b}$ 投影在 $\mathbf{a}$ 上的数值是相等的 ...