系统与基础设施

• What Every Developer Should Know About GPU Computing freeDone with What Every Developer Should Know About GPU Computing

  article · 40 分钟 — 一篇把 GPU 的硬件形状（SM、warp、寄存器与显存的层级）讲到应用工程师够用的入门，读完能说清 GPU 为什么吞得下矩阵乘、为什么数据搬运才是瓶颈。

• Making Deep Learning Go Brrrr From First Principles freeDone with Making Deep Learning Go Brrrr From First Principles

  article · 45 分钟 — 三种瓶颈（算力、带宽、开销）的判断法，连带算术强度与 FLOPs 估算：面试里被问「这个模型跑慢了卡在哪」，答案的框架就是这篇。

• How is LLaMa.cpp possible? freeDone with How is LLaMa.cpp possible?

  article · 30 分钟 — 把前两篇的原理套到 LLM 解码上算一遍：权重字节数 ÷ 带宽 = 每 token 的下限，于是量化为什么直接换来速度、一台笔记本为什么跑得动 7B，数字都自己算得出来。

• LLM Inference Handbook — Modular freeDone with LLM Inference Handbook — Modular

  docs — Latency, throughput and cost of serving a model — batching, KV cache, quantization, which engine when — as one practical handbook rather than twenty blog posts.

• Inside vLLM: Anatomy of a High-Throughput LLM Inference System freeDone with Inside vLLM: Anatomy of a High-Throughput LLM Inference System

  article · 90 min — One request followed through the scheduler, paged KV cache, continuous batching, prefix caching and speculative decoding, ending on the latency-against-throughput curve a design answer needs.

• vLLM documentation freeDone with vLLM documentation

  docs — The engine the cloud-GPU route runs on: the docs explain the trade-off behind each setting, and the bundled vllm bench serve prints TTFT, TPOT and throughput — the deployment project's numbers.

复现项目: Serve a small open model and measure it — Serve a small open model with vLLM, then measure time to first token and tokens per second at 1, 8 and 32 concurrent requests, with and without quantization.

• The Ultra-Scale Playbook: Training LLMs on GPU Clusters freeDone with The Ultra-Scale Playbook: Training LLMs on GPU Clusters

  book · 一个周末 — 从单卡显存算起，一步步引入数据并行、ZeRO、张量并行、流水线并行、上下文并行与专家并行，每一步都说明它省了什么、新增了哪种通信；读完能解释为什么训练要上千卡。

• How to Scale Your Model: A Systems View of LLMs on TPUs freeDone with How to Scale Your Model: A Systems View of LLMs on TPUs

  book · 一个周末 — 用 roofline 一把尺子贯穿训练与推理：每种并行策略的通信量和计算量都写成公式，让你能在白板上推导某个模型用多少芯片、什么切法最划算；第 12 章把同一套方法搬到 GPU。

• CS336: Language Modeling from Scratch freeDone with CS336: Language Modeling from Scratch

  course · 系统部分 4 讲约 6 小时 — 想听课而不是读书的那条路：第 5 讲 GPU、第 7–8 讲并行、第 10 讲推理，由一线研究者讲清系统设计的取舍，正好覆盖本层四个节点；讲座视频在 YouTube 播放列表里。

• DeepSeek-V3/R1 Inference System Overview freeDone with DeepSeek-V3/R1 Inference System Overview

  article · 30 分钟 — 一家模型公司亲手公开的推理账本：多少节点、每天处理多少 token、每天的 GPU 成本和按标价算出的收入，是「每 token 成本怎么算」唯一带真实数字的一手范例，连带解释了为什么大 batch 和专家并行能把成本压下去。

• Memory has grown to nearly two-thirds of AI chip component costs freeDone with Memory has grown to nearly two-thirds of AI chip component costs

  article · 15 分钟 — 供给侧的一个关键事实：一块 AI 芯片的成本大头已经是 HBM 而不是逻辑芯片，由此能解释为什么显存带宽贵、为什么算力扩张卡在内存供应而不只是晶圆。

• Learn Kubernetes Basics — Kubernetes documentation freeDone with Learn Kubernetes Basics — Kubernetes documentation

  docs — The project's own six-module tutorial: create a cluster, deploy, expose, scale and roll out an app — the vocabulary every customer's environment is written in, in an afternoon.

• Troubleshooting Applications — Kubernetes documentation freeDone with Troubleshooting Applications — Kubernetes documentation

  docs — Pods, services and running containers debugged with what you get on a cluster you do not own — describe, logs, events, a shell — the job when a customer's deployment misbehaves.

复现项目: Deploy your agent where a customer would run it — Package the agent as a container, deploy it to a local cluster with its secrets, a health probe and resource limits, then kill a node and watch what happens.