<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>项目 on Yiwen Cai</title><link>https://yiwen-cai.github.io/projects/</link><description>Recent content in 项目 on Yiwen Cai</description><generator>Hugo -- gohugo.io</generator><language>zh-cn</language><managingEditor>caiyiwen.cs@foxmail.com (Yiwen Cai)</managingEditor><webMaster>caiyiwen.cs@foxmail.com (Yiwen Cai)</webMaster><copyright>© 2026 Yiwen Cai</copyright><lastBuildDate>Sun, 12 Jul 2026 22:41:45 +0800</lastBuildDate><atom:link href="https://yiwen-cai.github.io/projects/index.xml" rel="self" type="application/rss+xml"/><item><title>kernel-skills</title><link/><pubDate>Sat, 11 Jul 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid/><description>面向 Claude Code / Codex 的 GPU kernel 开发 skill 集合，按 CUDA / Triton / NPU 分目录组织。已实装 CUTLASS/CuTe 环境配置（依赖检测、编译模板、VSCode IntelliSense、故障排查）与算子精度对拍（数值对拍 SOP、多 dtype 容差查表、TF32 污染规避），适配 sm_120（Blackwell）。</description></item><item><title>FastGNN：TensorCore GNN 训练系统</title><link/><pubDate>Fri, 01 May 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid/><description>H100 上基于 TensorCore 的图神经网络训练系统，自研稀疏算子后端（Voltrix block-sparse SpMM、Fused3S 稀疏注意力、Rabbit 节点重排），相对 DGL/PyG 1.8–3.3× 加速。</description></item><item><title>Fused Add + RMSNorm Triton Kernel</title><link/><pubDate>Wed, 01 Apr 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid/><description>用 Triton 融合 Add 与 RMSNorm，多行 program 映射 + hidden_size 分档调度，相比 PyTorch 基线 19.28× 加速。</description></item><item><title>xLLM 推理引擎适配（Qwen3.5）</title><link/><pubDate>Wed, 01 Apr 2026 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid/><description>基于 xllm 推理引擎适配 Qwen3.5，实现 GDN 线性注意力与 MTP 投机采样，解决 KV cache 异常，跑通 32k 输入推理。</description></item><item><title>CUDA GEMM Kernels</title><link/><pubDate>Sun, 01 Jun 2025 00:00:00 +0000</pubDate><author>caiyiwen.cs@foxmail.com (Yiwen Cai)</author><guid/><description>从零手写 8 个 CUDA GEMM kernel，覆盖 FP32 CUDA Core 与 FP16 Tensor Core WMMA；SGEMM v3 达 cuBLAS 的 95.1%，HGEMM v3 达 213 TFLOPS。</description></item></channel></rss>