Blog

Readnest:我的阅读记录,为什么要搬出 NotionReadnest: Why I Moved My Reading Log Out of Notion

一个只存在于自己手机里的阅读记录本A reading log that lives on your phone and nowhere else

2026-10-10

My reading records were scattered across WeRead, Zhangyue, JD Read, my Boox device and library paperbacks, held together by a Notion table I maintained by ha...

交叉熵：语言模型的损失函数，为什么只能是它Cross-Entropy: The Loss Function of Language Models, and Why It Must Be This One

3Blue1Brown「压缩即智能」Part 2 学习笔记——从一个玩具编码方案讲交叉熵，为什么 p 和 q 不能对调，又为什么交叉熵公式是「平均损失仅在 q=p 时取最小值」的唯一解

2026-10-05

Notes on 3Blue1Brown's 'Compression is Intelligence' Part 2. From a toy code that sends four commands to a remote robot — probabilities 1/2, 1/4, 1/8, 1/8 co...

千卡训练的隐形账单：通信不是瓶颈The Hidden Bill of Thousand-GPU Training: Communication Is Not the Bottleneck

MLSys·im 学习笔记（五）——从单节点走到千卡集群：谁在偷走训练时间、Young-Daly 公式该往哪用、选址怎么在 61 倍里改写碳足迹MLSys·im Learning Notes (Part 5) — from one node to a thousand GPUs: who actually steals your training time, where the Young-Daly formula should point, and how site selection rewrites the carbon footprint by 61x

2026-09-23

MLSys 学习的第五阶段记录:基于 Datawhale mlsysim 打卡任务 Task5(集群、成本与整合),复现教程 06 的 8→1024 GPU 规模扫描,并把通信开销拆成 TP 与 DP 两个分量——发现教程表里'32 卡通信比 8 卡还少'不是通信变好,而是 batch 在单机阶段被灌到 64、跨...

千卡训练的隐形账单：通信不是瓶颈The Hidden Bill of Thousand-GPU Training: Communication Is Not the Bottleneck

MLSys·im 学习笔记（五）——从单节点走到千卡集群：谁在偷走训练时间、Young-Daly 公式该往哪用、选址怎么在 61 倍里改写碳足迹MLSys·im Learning Notes (Part 5) — from one node to a thousand GPUs: who actually steals your training time, where the Young-Daly formula should point, and how site selection rewrites the carbon footprint by 61x

2026-09-23

Fifth stage of my MLSys learning: following the Datawhale mlsysim task (Cluster, Cost & Integration), I reproduce Tutorial 06's 8→1024 GPU scale sweep and sp...

给开源教程修了个 demoFixing a Tutorial Demo: An Open-Source PR Retrospective

给 mlsysim 的 KV-Cache 教程重写了 §5 的 Paged Attention demo 并提了 PRI rewrote the §5 Paged Attention demo in mlsysim's KV-Cache tutorial and opened PR #2135; before merging, the maintainer revised it and shrank my numbers from 314 to 259. A retrospective of that experience.

2026-09-16

一次开源贡献的完整复盘:读 mlsysim 的 KV-Cache 教程时发现 §5 Paged Attention demo 对三种页大小输出完全相同的结果——max_batch_size 绑定、seq_len 整除所有页大小、引擎只建模末页内部碎片、加速比是固定 0.6 因子。我重写了整个 demo(静态连续 ...

GPU 为什么会挨饿：数据墙、投机解码与设计空间探索Why Does the GPU Starve: The Data Wall, Speculative Decoding, and Design-Space Exploration

MLSys·im 学习笔记（四）——从单点瓶颈走向全链路优化：CPU 预处理如何让 GPU 饿肚子，投机解码如何以小博大，DSE 如何在数千种配置中自动找最优解MLSys·im Learning Notes (Part 4) — from single-point analysis to whole-pipeline optimization: how CPU preprocessing starves the GPU, how speculative decoding bets small to win big, and how DSE searches thousands of configs automatically

2026-09-16

Fourth stage of my MLSys learning: following the Datawhale mlsysim task (Data, Algorithms & System Optimization), I verify the data wall stage by stage — Res...

GPU 为什么会挨饿：数据墙、投机解码与设计空间探索Why Does the GPU Starve: The Data Wall, Speculative Decoding, and Design-Space Exploration

MLSys·im 学习笔记（四）——从单点瓶颈走向全链路优化：CPU 预处理如何让 GPU 饿肚子，投机解码如何以小博大，DSE 如何在数千种配置中自动找最优解MLSys·im Learning Notes (Part 4) — from single-point analysis to whole-pipeline optimization: how CPU preprocessing starves the GPU, how speculative decoding bets small to win big, and how DSE searches thousands of configs automatically

2026-09-16

MLSys 学习的第四阶段记录:基于 Datawhale mlsysim 打卡任务 Task4(数据、算法与系统优化),亲手验证了数据墙的三个环节——ResNet-50 在 A100 上纯 GPU 计算 13.97ms(18,323 img/s)是天花板,存储 I/O 只用到 PCIe 的 28.6%,而 8 个...

给开源教程修了个 demoFixing a Tutorial Demo: An Open-Source PR Retrospective

给 mlsysim 的 KV-Cache 教程重写了 §5 的 Paged Attention demo 并提了 PR #2135;合并前维护者又改了一版,把数字从 314 改小到 259。复盘这次经历。I rewrote the §5 Paged Attention demo in mlsysim's KV-Cache tutorial and opened PR #2135; before merging, the maintainer revised it and shrank my numbers from 314 to 259. A retrospective of that experience.

2026-09-16

A complete retrospective of one open-source contribution: while reading mlsysim's KV-Cache tutorial I found that the §5 Paged Attention demo printed identica...

KV-Cache 的显存账本：长上下文、并发与量化剪枝的取舍The KV-Cache Memory Ledger: Long Context, Concurrency, and the Quantization–Pruning Trade-off

MLSys·im 学习笔记（三）——用公式算清一个请求吃多少显存，序列长度如何决定并发上限，以及为什么"压缩"和"剪枝"在推理提速上完全是两回事MLSys·im Learning Notes (Part 3) — a formula for how much memory one request eats, why sequence length decides your concurrency ceiling, and why 'compression' and 'pruning' are two completely different stories for inference speedup

2026-09-12

Third stage of my MLSys learning: following the Datawhale mlsysim task (KV-Cache & model properties), I hand-compute the KV-Cache memory ledger — each term o...

KV-Cache 的显存账本：长上下文、并发与量化剪枝的取舍The KV-Cache Memory Ledger: Long Context, Concurrency, and the Quantization–Pruning Trade-off

MLSys·im 学习笔记（三）——用公式算清一个请求吃多少显存，序列长度如何决定并发上限，以及为什么"压缩"和"剪枝"在推理提速上完全是两回事MLSys·im Learning Notes (Part 3) — a formula for how much memory one request eats, why sequence length decides your concurrency ceiling, and why 'compression' and 'pruning' are two completely different stories for inference speedup

2026-09-12

MLSys 学习的第三阶段记录:基于 Datawhale mlsysim 打卡任务 Task3(KV Cache 与模型属性),亲手算清 KV-Cache 的显存账本——公式每一项对应 Llama-3-8B 的哪个超参数,实测序列长度翻倍则并发上限减半(2K 260 → 4K 130 → 32K 16 → 128...

内存墙与两阶段：为什么 3.2 倍算力只换来 1.6 倍加速The Memory Wall & Two Phases: Why 3.2x More FLOPS Buys Only 1.6x Speedup

MLSys·im 学习笔记（二）——从 Memory Wall 到 Prefill/Decode 双墙，一次 LLM 请求里的两套瓶颈MLSys·im Learning Notes (Part 2) — from the Memory Wall to the Prefill/Decode dual walls: two bottlenecks inside a single LLM request

2026-09-09

MLSys 学习的第二阶段记录:基于 MLSysBook 的 mlsysim 教程 01 Memory Wall 与 02 Two Phases,亲手验证"升级 H100 算力 3.2 倍、LLM 推理却只快 1.6 倍"的内存墙现象,拆解 Prefill(TTFT)撞算力墙、Decode(ITL)撞内存墙的两阶...

内存墙与两阶段：为什么 3.2 倍算力只换来 1.6 倍加速The Memory Wall & Two Phases: Why 3.2x More FLOPS Buys Only 1.6x Speedup

MLSys·im 学习笔记（二）——从 Memory Wall 到 Prefill/Decode 双墙，一次 LLM 请求里的两套瓶颈MLSys·im Learning Notes (Part 2) — from the Memory Wall to the Prefill/Decode dual walls: two bottlenecks inside a single LLM request

2026-09-09

Second stage of my MLSys learning: following the MLSysBook mlsysim tutorials 'The Memory Wall' and 'Two Phases, One Request', I verify hands-on why upgrading...

用 5 行代码预测模型瓶颈：MLSys·im Roofline 教程学习笔记Predicting Model Bottlenecks in 5 Lines of Code: MLSys·im Roofline Tutorial Notes

从 Datawhale 打卡任务开始的性能分析之旅——ResNet-50、Llama-3 在 A100/H100 上的 roofline 实战与踩坑复盘MLSys·im Learning Notes (Part 1) — a performance-analysis journey from a Datawhale assignment: rooflining ResNet-50 and Llama-3 on A100/H100, with a post-mortem of every pitfall

2026-09-08

Two days of MLSys learning notes: following the MLSysBook mlsysim tutorial, hands-on Roofline analysis — using arithmetic intensity and the ridge point to ju...

用 5 行代码预测模型瓶颈：MLSys·im Roofline 教程学习笔记Predicting Model Bottlenecks in 5 Lines of Code: MLSys·im Roofline Tutorial Notes

从 Datawhale 打卡任务开始的性能分析之旅——ResNet-50、Llama-3 在 A100/H100 上的 roofline 实战与踩坑复盘A performance-analysis journey starting from a Datawhale assignment — rooflining ResNet-50 and Llama-3 on A100/H100, with a post-mortem of every pitfall

2026-09-08

这两天的 MLSys 学习记录:基于 MLSysBook 的 mlsysim 教程,亲手跑通 Roofline 模型分析——用算术强度与脊点判断 ResNet-50 在 A100/H100 上的瓶颈变迁,解析 LLM 推理的内存墙现象。文章包含完整的实验数据、学习过程中的六个疑惑与解答,以及我读引擎源码挖出的两个...

个人主页升级实战：Jekyll 增量改造的取舍与落地Upgrading My Personal Site: Jekyll Incremental Refactoring — Trade-offs and What Actually Shipped

从「要不要迁移框架」到「搜索、评论、SEO 全都要」的一次完整记录From 'should I migrate frameworks' to 'search, comments, SEO — I want it all'

2026-09-02

一次真实的个人主页升级实践：面对「要不要迁移到 Astro」的诱惑，我选择在现有 Jekyll 基础上做增量改造，最终落地了分类标签、全文搜索、Giscus 评论、全套 SEO、GA4 统计与打赏入口。本文记录完整的决策过程、实施细节与踩坑经验。

个人主页升级实战：Jekyll 增量改造的取舍与落地Upgrading My Personal Site: Jekyll Incremental Refactoring — Trade-offs and What Actually Shipped

从「要不要迁移框架」到「搜索、评论、SEO 全都要」的一次完整记录

2026-09-02

A real personal-site upgrade: resisting the temptation to migrate to Astro, I did incremental refactoring on the existing Jekyll setup and shipped categories...

傻瓜操作：GraphRAG、Ollama 本地部署及踩坑记录A Noob-Friendly Guide: Deploying GraphRAG with Local Ollama, and Every Pitfall I Hit

2024-08-28

GraphRAG 本地部署完整指南：从环境准备、依赖安装到项目初始化，一步步讲解如何将微软 GraphRAG 与本地 Ollama 模型结合，实现完全离线的知识库问答，并附上部署过程中的踩坑记录。

傻瓜操作：GraphRAG、Ollama 本地部署及踩坑记录A Noob-Friendly Guide: Deploying GraphRAG with Local Ollama, and Every Pitfall I Hit

2024-08-28

Complete GraphRAG local deployment guide: environment setup, dependency install, project init — wiring Microsoft GraphRAG to a local Ollama model for fully o...

学习最优化课程中的一些疑惑Questions from an Optimization Course: Cartesian Products and Convex Sets

2023-10-18

学习最优化课程时的几个疑问：卡氏积是什么？n 维空间中凸集的卡氏积含义？凸集的卡氏积是否仍然是凸集？附 GPT 讲解与证明过程。

学习最优化课程中的一些疑惑Questions from an Optimization Course: Cartesian Products and Convex Sets

2023-10-18

Questions raised while studying optimization: what is a Cartesian product, what does it mean for convex sets in n-dimensional space, and is the product of co...

文献笔记 A View of Cloud ComputingReading Notes: A View of Cloud Computing

2023-06-24

经典论文阅读笔记：Armbrust 等人对云计算的权威定义，以及影响云计算落地的 10 大障碍与对应的解决机会。

文献笔记 A View of Cloud ComputingReading Notes: A View of Cloud Computing

2023-06-24

Classic paper reading notes: Armbrust et al.'s authoritative definition of cloud computing and the 10 obstacles to adoption with their corresponding opportun...

文献笔记 A Brief Introduction to Distributed SystemsReading Notes: A Brief Introduction to Distributed Systems

2023-05-22

本文主要讨论了分布式系统。作者提供了关于分布式系统的概述，是一份很好的新手教程。 Fig.1 大纲 在论文的第一部分，作者简要回顾了计算机的演变，指出了分布式系统的两个技术基础：性能强大的微处理器的发展和高速计算机网络的发明。为什么这些技术很重要？ The result of these techn...

文献笔记 A Brief Introduction to Distributed SystemsReading Notes: A Brief Introduction to Distributed Systems

2023-05-22

Reading notes on Coulouris-style distributed systems overview: definition, transparency classifications, and common design goals.