LLM inference — Inicenoj

Archives

Readings

Projects

Publications

Links

About

3D Reconstruction

RL Foundations

ML Systems

HPC Games

Basics

Misc

神奇妙妙通信小问题

从 NCCL Ring LL 到 PCIe IPC AllReduce

KV Size on sm10+

对 Looped Transformer 的简要学习笔记

尝试对Kimi KDA的数学推导&算子实现分析

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G1: 一切的开始 (56→130)

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G2: 从算术指令到 Occupancy 的挣扎 (130→144)

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G3: Fragment Reshape 与架构级突破 (144→227)

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G4: SM103 + TMA + TCGen05 (227→221)

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G5: Pipeline 重构与双缓冲 (221→727)

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G7: 回归 1CTA (880→987)

Flash Attention CUDA Kernel 优化: 从 56 到 986 TFLOPS on B300 — G6: Cluster 与 2CTA 探索 (727→880)

TopK kernel 优化: from 20GB/s to 2024GB/s on B300

Paged attention kernel optimization(I)

MoE vs Dense Models in Inference

Streams and Concurrency on CUDA

Foundation of Reinforcement learning(V)

Foundation of Reinforcement learning(IV)

Foundation of Reinforcement learning(III)

Foundation of Reinforcement learning(II)

Foundation of Reinforcement learning(I)

nanoPD:一个 LLM P/D 分离推理引擎的实现笔记

3D Reconstruction Series

学习笔记：Tensor Parallelism（TP）

HPCGames 题解 D E 题

HPCGames 题解 A B C 题

SF3D 论文阅读记录

ViT Transformer 的阅读?(应该算是阅读吧)

回顾一下Transformer

SLAM Former 阅读

重返vggt

论文阅读记录：reloc3r

论文阅读记录：Fast3R

论文阅读记录：MAst3R

VGGT读后有感

为SLAM3R补充实时处理函数方法

SLAM3R读后有感

Celebrate and Introduce My First Page