刘芷溢
最近更新
FlashAttention v1-v4：从 IO 感知到异步流水与硬件协同
沿着 FlashAttention v1-v4 的论文、作者博客与源码，拆解 Ampere、Hopper、Blackwell 如何改变 attention 的数据路径、线程分工和瓶颈，并用调度图解释 copy、MMA、softmax 与输出校正如何真正重叠。
理解 H100：从硬件数据路径到 Nsight Compute 瓶颈分析
沿着 kernel 的分配、指令发射、访存与异步矩阵计算，建立 H100 的硬件工作模型，并把 Nsight Compute 的吞吐、调度、事务和依赖指标组织成可验证的瓶颈诊断方法。
Bridging the Gap Between Promise and Performance for Microscaling FP4 Quantization 论文笔记
以格式、量化算法和 GPU Kernel 的跨层约束为主线，拆解 ICLR 2026 论文 Bridging the Gap Between Promise and Performance for Microscaling FP4 Quantization，解释旋转为何改善 MXFP4 却可能伤害 NVFP4、MR-GPTQ 如何修正这一矛盾，以及 QuTLASS 的真实收益与部署边界。
FP4、FP8 与 INT8/INT4：数值范围、块缩放与推理计算精度
从标量编码、共享 scale 与有效位宽出发，拆解 INT8、FP8、MXFP8、INT4、MXFP4、NVFP4 的表示范围，并沿 GEMM、GQA/MLA Attention 数据流解释 8 bit/4 bit 推理真正使用的乘法、累加、Softmax 与输出精度。
从 EAGLE 到 DFlash 2：SGLang 与 SpecForge 的投机解码训练和推理
沿 EAGLE、EAGLE-2、EAGLE-3、DFlash、DSpark 和 DFlash 2 的演进，解释 draft 的训练目标、数据对齐、推理状态与验证预算，并对照 SpecForge 训练和 SGLang serving 源码串起完整流程。
README
CuTe FlashAttention-2 case-study benchmark This directory preserves the exact teaching kernel, benchmark harness, raw results, and figure generator used by the corresponding tom-jerr blog post.
CuTe 详解：以 FlashAttention-2 拆解 Layout、TiledCopy 与 TiledMMA
以一份可运行的 FlashAttention-2 前向 Kernel 为贯穿案例，从坐标函数和线程所有权出发，详解 CuTe 的 Layout algebra、Tensor、TiledCopy、TiledMMA、寄存器重解释与流水线，并用正确性和性能实验检查这些抽象最终生成了什么。
LLM 量化综述：数据格式、PTQ 与 QAT 的算法演进
从 INT8、FP8、INT4、MXFP4、NVFP4 的表示范围与舍入误差出发，按 8 bit、4 bit 和低于 4 bit 梳理 PTQ/QAT，解释各方法解决的问题、相对前作的创新、核心优化与数值稳定性，并对应到 llm-compressor 的工程实现。
vector search
5 篇文章
PANNS: Enhancing Graph-based Approximate Nearest Neighbor Search through Recency-aware Construction and Parameterized Search
PANNS: Enhancing Graph-based Approximate Nearest Neighbor Search through Recency-aware Construction and Parameterized Search. It analyzes the graph-based ANNS workload, offers a parameterized search strategy for flexible speed-accuracy trade-offs, and incorporates hidden dimensions through a new proximity graph construction algorithm and graph memory layout optimization.
PDX: A Data Layout for Vector Similarity Search
PDX: A Data Layout for Vector Similarity Search. The design of PDX, a new data layout for vectors alongside PDXearch: a framework to perform pruned VSS dimension-by-dimension. The design and evaluation of PDX-BOND leverages the PDX layout to visit first the most relevant dimensions relative to the incoming query. To incorporate hidden dimensions not embedded into vectors, it proposes a new proximity graph construction algorithm and a graph memory layout optimization.
iQAN: Fast and Accurate Vector Search with Efficient Intra-Query Parallelism on Multi-Core Architectures 论文阅读笔记
iQAN: Fast and Accurate Vector Search with Efficient Intra-Query Parallelism on Multi-Core Architectures. It studies the root causes of poor scalability in vector search on multi-core architectures and introduces path-wise parallelism, staged expansion, and redundancy-aware synchronization.
A Large-Scale Disk-Based System for Dynamic Vector Search 论文阅读笔记
LSM-VEC uses a write-optimized LSM-tree for graph indexes, keeps only the bottom layer on disk, applies locality-aware graph reordering, and uses sampling-guided traversal with probabilistic routing.
MIRAGE-ANNS: Mixed Approach Graph-based Indexing for Approximate Nearest Neighbor Searcha Layout for Vector Similarity Search
MIRAGE-ANNS: Mixed Approach Graph-based Indexing for Approximate Nearest Neighbor Searcha Layout for Vector Similarity Search. It constructs the index as fast as refinement-based approaches while retaining search performance comparable to or better than increment-based ones.
