记录 AI 系统、底层计算与工程实践的个人知识库，连接论文、源码、trace 与性能建模。https://weigao.cc/zh-CNhttps://weigao.cc/ai-systems/llm-inference/hyperloom-specialist-config-tuning/https://weigao.cc/ai-systems/llm-inference/hyperloom-specialist-config-tuning/拆解 AMD Hyperloom 的 specialist 实现与服务配置调优链路：knob 面、变体指纹、贪心爬山、衰减接受线与单次测量的方法学边界Wed, 05 Aug 2026 00:00:00 GMTllm-inferencegpu-optimizationagentichyperparameter-tuningbenchmarkrocmsglangvllmhttps://weigao.cc/ai-systems/llm-inference/vllm-async-scheduling-speculative-decoding/https://weigao.cc/ai-systems/llm-inference/vllm-async-scheduling-speculative-decoding/沿 vLLM 源码解释 async scheduling 的三态配置、output placeholder、EAGLE/MTP 状态归属，以及 accepted-count 同步如何影响真实重叠Wed, 05 Aug 2026 00:00:00 GMTllm-inferencevllmdecodekv-cachetracehttps://weigao.cc/ai-systems/llm-inference/kimi-k3-architecture-training-inference/https://weigao.cc/ai-systems/llm-inference/kimi-k3-architecture-training-inference/沿一次 Kimi K3 前向过程，系统解释 Hybrid KDA–MLA、AttnRes、Stable LatentMoE、长上下文训练和在线推理系统Tue, 04 Aug 2026 00:00:00 GMTllm-inferenceattentionmoedistributed-trainingquantizationhttps://weigao.cc/toolbox/tools/cmux-setup/https://weigao.cc/toolbox/tools/cmux-setup/cmux（Ghostty 内核的 macOS AI 终端）的配色、侧栏、远程 SSH 全套配置，附一次实配过程中踩到的 12 个坑与实测数据。Sat, 01 Aug 2026 00:00:00 GMTcmuxterminaldotfilesmacossshhttps://weigao.cc/ai-systems/llm-inference/inference-parallelism/https://weigao.cc/ai-systems/llm-inference/inference-parallelism/从显存、计算、通信与拓扑四个约束理解推理侧 DP、TP、PP、EP、CP 的作用域和组合方法Tue, 28 Jul 2026 00:00:00 GMTllm-inferenceparallelismtensor-parallelpipeline-parallelexpert-parallelcontext-parallelhttps://weigao.cc/ai-systems/llm-inference/kernel-runtime-optimization/https://weigao.cc/ai-systems/llm-inference/kernel-runtime-optimization/用 HBM 流量、并行度、Kernel launch 与同步四个控制点理解 FlashAttention、Decode Kernel、Fusion 和 CUDA GraphTue, 28 Jul 2026 00:00:00 GMTllm-inferencegpu-kernelflashattentionflashdecodekernel-fusioncuda-graphhttps://weigao.cc/ai-systems/llm-inference/kv-cache-hit-ratio-tpm-correction-model/https://weigao.cc/ai-systems/llm-inference/kv-cache-hit-ratio-tpm-correction-model/在固定配置下把 KV prefix hit 映射到 compute、cache load、overlap、TTFT 与 Delivered/Physical TPMTue, 28 Jul 2026 00:00:00 GMTllm-inferencekv-cachesimulatorprefilltpmhttps://weigao.cc/ai-systems/llm-inference/moe-inference/https://weigao.cc/ai-systems/llm-inference/moe-inference/MoE 推理中的参数口径、dispatch-compute-combine 数据流、EP 边界与显存合同Tue, 28 Jul 2026 00:00:00 GMTmoeexpert-parallelisminferencehttps://weigao.cc/ai-systems/llm-inference/simulator-modeling-guide/https://weigao.cc/ai-systems/llm-inference/simulator-modeling-guide/LLM 推理容量模拟器的场景输入、显存、吞吐、scope、低精度格式与校准合同Tue, 28 Jul 2026 00:00:00 GMTsimulatormemory-modelinginferencehttps://weigao.cc/ai-systems/llm-inference/token-flow-hidden-state/https://weigao.cc/ai-systems/llm-inference/token-flow-hidden-state/用一次 decode step 串起 hidden state、Attention、FFN/MoE、LM Head 和 SamplingTue, 28 Jul 2026 00:00:00 GMTllm-inferencetransformerhidden-stateattentionmoelm-headhttps://weigao.cc/ai-systems/llm-inference/00-full-stack-synthesis/https://weigao.cc/ai-systems/llm-inference/00-full-stack-synthesis/用算力、带宽、容量与通信四种压力串起 Token 数据流、KV Cache、量化、调度、并行、Kernel 和 ServingSun, 26 Jul 2026 00:00:00 GMTllm-inferencesynthesissystemsoptimizationhttps://weigao.cc/ai-systems/llm-inference/01-kv-cache/https://weigao.cc/ai-systems/llm-inference/01-kv-cache/理解 KV Cache 的复用语义、容量公式、分页与共享机制，以及从分配到驱逐的生命周期Sun, 26 Jul 2026 00:00:00 GMTLLMInferenceKV CachePagedAttentionvLLMMemory Managementhttps://weigao.cc/ai-systems/llm-inference/03-quantization/https://weigao.cc/ai-systems/llm-inference/03-quantization/从第一性原理解释量化对象、数值格式、Scale 粒度、PTQ/QAT、运行时 Kernel 与性能验收边界Sun, 26 Jul 2026 00:00:00 GMTllm-inferencequantizationprefilldecodegpuhttps://weigao.cc/ai-systems/llm-inference/cpu-engineer-quantization-bridge/https://weigao.cc/ai-systems/llm-inference/cpu-engineer-quantization-bridge/把 INT8、VNNI、AMX、Cache、Memory Wall 和 perf/IBS 的 CPU 心智模型映射到 Scale、Tensor Core、HBM 与 GPU Kernel 验证Sun, 26 Jul 2026 00:00:00 GMTllm-inferencequantizationcpugpuhttps://weigao.cc/ai-systems/llm-inference/fp4-fp8-quantization/https://weigao.cc/ai-systems/llm-inference/fp4-fp8-quantization/区分 FP4/FP8 数值格式、Current/Delayed/Block scaling、NVFP4 与 MXFP8，并说明 Checkpoint 到 Kernel 的能力边界Sun, 26 Jul 2026 00:00:00 GMTllm-inferencequantizationgpuhttps://weigao.cc/ai-systems/llm-inference/glm52-operator-quantization/https://weigao.cc/ai-systems/llm-inference/glm52-operator-quantization/用待 Trace 复核的 GLM-5.2 Prefill 示例路径解释 W4A8、E2M1/E4M3、低精度 BMM、BF16 Attention 与 FP8 KV CacheSun, 26 Jul 2026 00:00:00 GMTllm-inferencequantizationfp4fp8nvfp4profilinghttps://weigao.cc/ai-systems/llm-inference/quantization-methods-evaluation/https://weigao.cc/ai-systems/llm-inference/quantization-methods-evaluation/按控制点理解 GPTQ、AWQ、SmoothQuant、Rotation 与 QAT，并用执行真实性、质量、容量、性能四道门验收量化方案Sun, 26 Jul 2026 00:00:00 GMTllm-inferencequantizationprofilingresearchhttps://weigao.cc/cpu-gpu/cpu-architecture/instruction-sets/amx/https://weigao.cc/cpu-gpu/cpu-architecture/instruction-sets/amx/Intel AMX 的 Tile 寄存器、TMUL、低精度输入与宽精度累加合同Sun, 26 Jul 2026 00:00:00 GMTx86architectureComputer Architecturehttps://weigao.cc/ai-systems/llm-inference/attention-evolution/https://weigao.cc/ai-systems/llm-inference/attention-evolution/用 KV 共享、表示压缩、访问稀疏和递推状态四条路线理解现代 LLM AttentionTue, 21 Jul 2026 00:00:00 GMTllm-inferenceattentionkv-cachehttps://weigao.cc/ai-systems/profiling/profiling-to-simulation-evidence-chain/https://weigao.cc/ai-systems/profiling/profiling-to-simulation-evidence-chain/把线上 profiling 证据和仿真假设验证组织到同一个推理性能分析框架里Sat, 11 Jul 2026 00:00:00 GMTprofilingsimulationllm-inferenceperformance-analysisawphttps://weigao.cc/ai-systems/llm-inference/prefill-trace-worker-dsa-mla/https://weigao.cc/ai-systems/llm-inference/prefill-trace-worker-dsa-mla/用 GLM-5.2 Prefill Trace 区分 worker 供给、模型计算、KV 数据路径、通信与同步Fri, 10 Jul 2026 00:00:00 GMTllm-inferenceprefilltracemlachunked-prefillschedulinghttps://weigao.cc/ai-systems/llm-inference/gdn-chunked-prefill/https://weigao.cc/ai-systems/llm-inference/gdn-chunked-prefill/区分 scheduler chunk 与 GDN kernel chunk，并解释 prepare_chunk_indices 的映射和同步边界Thu, 09 Jul 2026 00:00:00 GMTllm-inferencegdnqwen3nextchunked-prefillprefillvllmprofilinghttps://weigao.cc/ai-systems/llm-inference/deepseek-mla/https://weigao.cc/ai-systems/llm-inference/deepseek-mla/解释 MLA 的低秩 KV、RoPE 解耦、矩阵吸收、容量公式与 serving 边界Mon, 06 Jul 2026 00:00:00 GMTllm-inferenceattentionkv-cachedeepseekmlahttps://weigao.cc/ai-systems/llm-inference/chunked-prefill-deep-dive/https://weigao.cc/ai-systems/llm-inference/chunked-prefill-deep-dive/从调度合同、Attention/GEMM 形状与验收矩阵理解 Chunked Prefill 的收益和边界Thu, 02 Jul 2026 00:00:00 GMTllm-inferencechunked-prefillschedulingprefilldecodeattentionvllmhttps://weigao.cc/ai-systems/llm-inference/inference-frameworks-2026/https://weigao.cc/ai-systems/llm-inference/inference-frameworks-2026/用 Serving Stack、实现差异和选型合同比较 LLM 推理框架，而不是寻找脱离 workload 的最快引擎Wed, 01 Jul 2026 00:00:00 GMTvllmsglangtensorrt-llminference-frameworkdisaggregated-servingkv-cachehttps://weigao.cc/ai-systems/llm-inference/deepseek-v3-technical-report-bilingual-notes/https://weigao.cc/ai-systems/llm-inference/deepseek-v3-technical-report-bilingual-notes/对 DeepSeek-V3 Technical Report 的中英对照导读：架构、MoE 负载均衡、MTP、FP8 训练、Prefill/Decode 部署和推理工程启发Mon, 29 Jun 2026 00:00:00 GMTLLMDeepSeekMoEFP8MTPInferencehttps://weigao.cc/ai-systems/llm-inference/dspark-vs-mtp/https://weigao.cc/ai-systems/llm-inference/dspark-vs-mtp/调研 DeepSeek DSpark 的半自回归草稿模型、置信度调度验证、DeepSpec 源码实现，并和 MTP-1 对比其工程边界Mon, 29 Jun 2026 00:00:00 GMTllm-inferencespeculative-decodingdeepseekdsparkmtphttps://weigao.cc/toolbox/tools/html-diagram-test/https://weigao.cc/toolbox/tools/html-diagram-test/验证文章中嵌入 agent 生成的 HTML 交互图。Wed, 17 Jun 2026 00:00:00 GMThtml-diagraminteractivehttps://weigao.cc/ai-systems/llm-inference/causal-attention-kv-hit-area/https://weigao.cc/ai-systems/llm-inference/causal-attention-kv-hit-area/用 causal attention 的下三角区域推导 prefix cache hit 后 attention 近似按 1-h² 缩放Mon, 15 Jun 2026 00:00:00 GMTllm-inferenceattentionkv-cachesimulatorprefillhttps://weigao.cc/ai-systems/llm-inference/csa-hca-attention/https://weigao.cc/ai-systems/llm-inference/csa-hca-attention/基于摄入材料整理 CSA/HCA 的逐层压缩、Indexer、异构 cache 与百万上下文估算，并明确待核验边界Mon, 01 Jun 2026 00:00:00 GMTattentionsparse-attentionkv-cachedeepseekhttps://weigao.cc/toolbox/awp-profiling-api/https://weigao.cc/toolbox/awp-profiling-api/AWP 性能分析平台的 REST API 文档：异步提交 GPU/CPU profiling 任务并查询分析结果Mon, 01 Jun 2026 00:00:00 GMTprofilingapigpucputoolingawphttps://weigao.cc/toolbox/tools/zsh-setup/https://weigao.cc/toolbox/tools/zsh-setup/一份可一键复刻的 Zsh / Oh-My-Zsh 配置：插件、主题、别名与跨机迁移脚本。Sat, 30 May 2026 00:00:00 GMTzshdotfilesmacoshttps://weigao.cc/fundamentals/algorithms/dfs_bfs/https://weigao.cc/fundamentals/algorithms/dfs_bfs/深度优先搜索与广度优先搜索的原理及应用Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/algorithms/dp/https://weigao.cc/fundamentals/algorithms/dp/动态规划基础：斐波那契、记忆化搜索与状态转移Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/algorithms/presum/https://weigao.cc/fundamentals/algorithms/presum/前缀和算法原理与经典例题汇总Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/algorithms/slide_window/https://weigao.cc/fundamentals/algorithms/slide_window/滑动窗口与双指针技巧及例题Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/algorithms/sort/https://weigao.cc/fundamentals/algorithms/sort/排序算法与 Python 自定义排序技巧Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/effective-python/https://weigao.cc/fundamentals/programming-languages/python/language/effective-python/Effective Python 编程技巧与最佳实践Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/log/https://weigao.cc/fundamentals/programming-languages/python/language/log/Python logging 模块使用与日志级别配置Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/pip/https://weigao.cc/fundamentals/programming-languages/python/language/pip/pip 镜像源配置与包管理技巧Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/py-data-struct/https://weigao.cc/fundamentals/programming-languages/python/language/py-data-struct/Python 数据结构：切片、列表与元组操作Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/pytest/https://weigao.cc/fundamentals/programming-languages/python/language/pytest/Pytest 与 Unittest 测试框架使用指南Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/python_c/https://weigao.cc/fundamentals/programming-languages/python/language/python_c/Python 与 C/C++ 混合编程方法Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/language/python-function/https://weigao.cc/fundamentals/programming-languages/python/language/python-function/Python 高阶函数与字符串处理技巧Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/fundamentals/programming-languages/python/tools/py-tools/https://weigao.cc/fundamentals/programming-languages/python/tools/py-tools/IPython、timeit 等 Python 开发辅助工具Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/linux/system-programming/c-pointer/https://weigao.cc/linux/system-programming/c-pointer/C 语言指针的定义、用法与常见陷阱Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/linux/system-programming/cmake/https://weigao.cc/linux/system-programming/cmake/CMake 构建系统的核心概念与常用配置Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/linux/system-programming/cmake_makefile/https://weigao.cc/linux/system-programming/cmake_makefile/Makefile 规则语法与编写实践Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/linux/system-programming/size_t/https://weigao.cc/linux/system-programming/size_t/C/C++ 中 size_t 类型的定义、原理与使用场景Tue, 26 May 2026 00:00:00 GMThttps://weigao.cc/toolbox/research-projects/papers/vp_hpca14/https://weigao.cc/toolbox/research-projects/papers/vp_hpca14/HPCA14 论文解读：上下文相关的 Load Value 预测器Tue, 26 May 2026 00:00:00 GMT