Extensive Reading A similar paper is found: arxiv.org/pdf/2504.08791? Author Info ‪Zonghang Li‬ - ‪Google 学术搜索‬ Background LLM serving is shifting from the cloud to edge devices like smartphones and laptops. This trend is driven by growing privacy concerns, as users want to avoid sending their sensitive interaction data to cloud providers. The goal is to process user requests locally on their own devices. Preliminaries TP 场景下 KV Cache 的维护 Challenges Hardware Limitations: Mobile devices have very limited memory (typically 4-16 GiB) and computing power, often lacking GPUs. Running a 70B-scale model can require over 40 GiB of memory, which far exceeds the capacity of a single device. Inefficient Parallelism: The standard solution for distributed systems, pipeline parallelism, is inefficient for home scenarios where only one request is processed at a time. This leads to many devices being idle most of the time, wasting resources. Slow Memory Offloading: Existing on-device solutions like llama.cpp and Accelerate offload model data to disk to save RAM. However, their blocking disk I/O operations significantly slow down the inference speed. Insights 在低资源设备协同的环境下，应该选择 Tensor Parallelism 用户的请求一次性只有一条，并行的目的应该是降低延迟而不是增加吞吐量 Tensor Parallelism 依赖 allreduce 操作来同步和聚合计算结果，在低资源设备协同的环境下，通信瓶颈并非网络带宽而是链路延迟 使用 star-based allreduce 来降低网络跳数进而降低延迟 使用滑动窗口内存调度器来异步加载和卸载权重 由一个独立的线程后台进行 将权重的加载隐藏在计算和同步的过程中 Approaches ...

Extensive Reading Author Info About Me — Tim Dettmers: A research scientist at the Allen Institute for Artificial Intelligence (Ai2) and an incoming Assistant Professor at Carnegie Mellon University (CMU). ‪Mike Lewis‬ - ‪Google Scholar‬ Related Blogs LLM.int8() and Emergent Features — Tim Dettmers A Gentle Introduction to 8-bit Matrix Multiplication for transformers at scale using transformers, accelerate and bitsandbytes Background 常见的 8 bit 量化有两种： Absmax quantization 在所有数据中，找到绝对值的最大值，我们称之为 abs_max，然后计算一个全局的 scaling factor 用数据中的每一个数字乘以这个 scaling factor, 再四舍五入到最近的整数完成量化 Zeropoint Quantization 找到数据中的最大值和最小值，计算 scaling factor 同时引入一个偏移量 zeropoint 来利用整个映射后的数值范围 精度更高，但是开销更大 Challenges How to preserve high quantization precision at scales beyond 1B parameters? How to deal with the systematic outliers emerged in all transformer layers starting at scales of 6.7B parameters? Insights Regular quantization methods introduce larger quantization errors for outliers. The amount of outlier can be small, but contributes the majority to the LLM’s quality. Isolate the outlier feature dimensions into a 16-bit matrix multiplication while other values are multiplied in 8-bit. Approaches 主要包括两部分内容： ...

Q: 能否用一个更详细的例子说明 git worktree 好的，没问题。让我们通过一个完整、详细的实际场景来演练一下 git worktree 的用法。 场景设定 你的项目：一个名为 my-website 的网站项目。 你的主目录：~/dev/my-website 你的主要任务：你正在开发一个复杂的新功能——“图片画廊”（Photo Gallery）。为此，你创建了一个名为 feature/gallery 的分支，并且已经写了一些代码，但工作远未完成。 突发状况：产品经理跑过来说，网站页脚的版权年份还是去年，这是一个紧急的线上 Bug，必须马上修复！ 现在，你的 feature/gallery 分支工作区是“凌乱”的，有很多新建的、修改过的但还不能提交的文件。 详细步骤演练 第 1 步：检查当前状态 首先，我们看看当前的工作目录。 # You are in your main project directory cd ~/dev/my-website # You are working on the gallery feature git status git status 的输出可能是这样的： On branch feature/gallery Changes not staged for commit: (use "git add <file>..." to update what will be committed) (use "git restore <file>..." to discard changes in working directory) modified: src/app.js Untracked files: (use "git add <file>..." to include in what will be committed) src/components/PhotoGallery.vue no changes added to commit (use "git add" and/or "git commit -a") 此时，你不能直接切换到 main 分支，因为这会强制你处理这些未提交的修改，这正是我们想避免的。 ...

Intensive Reading Author Info Zichang Liu：Research Scientist at Meta. Jue Wang, Ph.D: Founder & President of Stylar AI (stylar.ai). Tri Dao: Assistant Professor of Computer Science at Princeton University. Chief Scientist at Together AI. Background LLM Inference Latency Breakdown Challenges Speeding up inference-time sparse LLMs in wall-clock time while maintaining quality and in-context learning abilities remains a challenging problem. While sparsity and pruning have been well-studied, they have not seen wide adoption on LLMs due to the poor quality and efficiency trade-offs on modern hardware such as GPUs: ...

Math Vector-Matrix Multiplication 从三个不同的角度分析向量乘以矩阵的运算过程 $xW$。 假设向量 $x$ 的形状是 $(1, 3)$，矩阵 $W$ 的形状是 $(3, 6)$。 $$x = \begin{bmatrix} x_1 & x_2 & x_3 \end{bmatrix}$$$$ W = \begin{bmatrix} w_{11} & w_{12} & w_{13} & w_{14} & w_{15} & w_{16} \\\\ w_{21} & w_{22} & w_{23} & w_{24} & w_{25} & w_{26} \\\\ w_{31} & w_{32} & w_{33} & w_{34} & w_{35} & w_{36} \end{bmatrix} $$根据矩阵乘法规则，结果 $y = xW$ 的形状将是 $(1, 6)$。 角度一：将 W 视为元素的二维集合 这是最基本、最微观的视角。我们将矩阵 $W$ 看作是一个 $3 \times 6$ 的数字网格。结果向量 $y$ 中的每一个元素 $y_j$，都是通过将向量 $x$ 的每个元素与其在矩阵 $W$ 中对应列的每个元素相乘，然后将结果相加得到的。 ...

Intensive Reading Author Info ‪Keivan Alizadeh-Vahid‬ - ‪Google Scholar‬ Iman Mirzadeh: An ML Research Engineer at Apple. Background LLM is hard for personal devices to load. The standard approach is to load the entire model into DRAM (Dynamic Random Access Memory) for inference. However, this severely limits the maximum model size that can be run. Challenges The primary challenge is that the memory footprint of large language models (LLMs) often exceeds the limited DRAM capacity of personal devices. While storing models on high-capacity flash memory is a potential solution, it introduces two new major challenges: ...

Intensive Reading Author Info Zhenliang Xue: From IPADS. Yixin Song: First author of PowerInfer. Zeyu Mi (糜泽羽): He is an associate professor at School of Software, Shanghai Jiao Tong University (SJTU). Haibo Chen [IPADS]: Director of Institute of Parallel and Distributed Systems. Background Sparsity FFN 的参数占比大，稀疏化特征也明显（特别是在使用 ReLU 时），所以可以在执行计算前利用一个 predictor 来预测哪些神经元会被激活，从而降低计算和 I/O 开销。 PowerInfer2 还探索了 LLM 推理过程中的动态稀疏性： 当批次很大时，对于任何一个神经元，只要它被输入中的至少一个激活，它在这一步的计算中就不是稀疏的。由于不同输入会激活不同神经元，其聚合效应导致大量神经元被激活，形成稳定、密集的“热点”，整体稀疏度显著降低。 由于某些序列会更早终止，所以有效批次的大小也会动态波动。这个实时变化导致了模型的计算模式在一个任务的生命周期内，会从一个接近稠密的模式平滑地过渡到一个高度稀疏的模式。 Mobile Hardware Characteristics 与 PC 相比，手机的硬件有两个特点： Heterogeneous computing capabilities with distinct sparse computation characteristics. CPU 更擅长稀疏计算 NPU 更擅长稠密计算 GPU 比 CPU 和 NPU 都更慢，而且在推理中使用 GPU 会影响设备的渲染帧率 移动 LLM 推理框架应同时利用异构处理器，以最大限度地利用共享内存带宽 Distinct storage architecture with unique I/O characteristics. 读的块大小越大，吞吐量越高 数据范围越小，吞吐量越高 频率越高的 CPU core 读取时吞吐量越高 UFS 并发能力有限 ...

Extensive Reading Author Info Ji Lin’s Homepage Jiaming Tang Shang Yang | MIT EECS Song Han - Associate Professor, MIT EECS Background Quantization is vital for running LLM on edge devices. Challenges Quantization-aware training (QAT) is not efficient due to the high training cost. Post-training quantization (PTQ) suffers from large accuracy degradation under a low-bit setting. Insights Not all weights in an LLM are equally important. Protecting only 1% salient weights can greatly reduce quantization error. To identify salient weight channels, we should refer to the activation distribution, not weights. Mixed-precision format is not hardware-efficient, we can employ activation-aware scaling. Approaches Activation-aware Weight Quantization ...

Intensive Reading Author Info ‪Yixin Song‬ - ‪Google Scholar‬ Zeyu Mi (糜泽羽): He is an associate professor at School of Software, Shanghai Jiao Tong University (SJTU). Haotong Xie (谢昊彤) Haibo Chen [IPADS]: Director of Institute of Parallel and Distributed Systems. Background Local deployments focus on low latency in processing small batches. LLM inference exhibits notable sparsity in neuron activation, a phenomenon observed in both self-attention and MLP blocks. The offloading technique leverages the CPU’s additional computational and memory resources. GPU-centric offloading utilizes CPU memory to store portions of the model parameters that exceed the GPU’s capacity. Lead to substantial per-token latency mainly due to frequent data transfers between GPU and CPU. Over 99.5% of processing time is consumed by transferring LLM weights from CPU to GPU. Hybrid offloading distributes model parameters between GPU and CPU, splitting them at the Transformer layer level. The CPU processes its layers first, then sends intermediate results to the GPU for token generation. The CPU, with higher memory but lower computational power, ends up handling 98% of the total computational time. ...

Extensive Reading Author Info Ying Sheng: She got her Ph.D. in Computer Science at Stanford University (Centaur), where she was advised by Clark Barrett. Before that, she received an M.S. in Computer Science from Columbia University in 2017 and a B.E. in Computer Science and Technology from ACM Honored Class, Shanghai Jiao Tong University in 2016. Lianmin Zheng: He is a member of technical staff at xAI. His research interests include machine learning systems, large language models, compilers, and distributed systems. Previously, he completed his Ph.D. at UC Berkeley, where he was advised by Ion Stoica and Joseph E. Gonzalez. Binhang Yuan(袁彬航) – Assistant Profossor@CSE HKUST: He is an assistant professor in the Department of Computer Science & Engineering (CSE), also affiliated with World Sustainable Development Institute, at the Hong Kong University of Science and Technology (HKUST). He is leading the Relaxed System Lab. Background Prior efforts to lower resource requirements of LLM inference correspond to three directions: ...

Extensive Reading Author Info About | Edward Hu: Edward Hu is a founding partner in a stealth AI company in Woodside, CA. He was a researcher at OpenAI and received his research training as a Ph.D. student advised by Yoshua Bengio, a recipient of the 2018 A.M. Turing Award. Before graduate school, Edward was a researcher at Microsoft, where he invented LoRA and μTransfer. Yelong Shen - Microsoft | AMiner Background The dominant paradigm in modern NLP is ...