<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>后端与架构 · Rice 的技术博客</title><link>https://zxing2004.github.io/blog/categories/%E5%90%8E%E7%AB%AF%E4%B8%8E%E6%9E%B6%E6%9E%84/</link><description>后端与架构栏目的最新文章，来自Rice 的技术博客</description><generator>Hugo</generator><language>zh-CN</language><managingEditor>Rice (ricead@qq.com)</managingEditor><lastBuildDate>Thu, 05 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://zxing2004.github.io/blog/categories/%E5%90%8E%E7%AB%AF%E4%B8%8E%E6%9E%B6%E6%9E%84/index.xml" rel="self" type="application/rss+xml"/><item><title>INT4 量化推理实战：精度补偿与吞吐提升的取舍</title><link>https://zxing2004.github.io/blog/posts/int4-%E9%87%8F%E5%8C%96%E6%8E%A8%E7%90%86%E5%AE%9E%E6%88%98%E7%B2%BE%E5%BA%A6%E8%A1%A5%E5%81%BF%E4%B8%8E%E5%90%9E%E5%90%90%E6%8F%90%E5%8D%87%E7%9A%84%E5%8F%96%E8%88%8D/</link><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/int4-%E9%87%8F%E5%8C%96%E6%8E%A8%E7%90%86%E5%AE%9E%E6%88%98%E7%B2%BE%E5%BA%A6%E8%A1%A5%E5%81%BF%E4%B8%8E%E5%90%9E%E5%90%90%E6%8F%90%E5%8D%87%E7%9A%84%E5%8F%96%E8%88%8D/</guid><description>INT4 全量化正在成为推理加速的主流选择。本文讲透线性量化原理、W4A4 的难度与精度补偿手段，并给出 PyTorch 量化实现和 vLLM/llama.cpp 部署命令。</description></item><item><title>开源模型的安全护栏：能力分级、Red Teaming 与许可证</title><link>https://zxing2004.github.io/blog/posts/%E5%BC%80%E6%BA%90%E6%A8%A1%E5%9E%8B%E7%9A%84%E5%AE%89%E5%85%A8%E6%8A%A4%E6%A0%8F%E8%83%BD%E5%8A%9B%E5%88%86%E7%BA%A7red-teaming-%E4%B8%8E%E8%AE%B8%E5%8F%AF%E8%AF%81/</link><pubDate>Sat, 10 Oct 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E5%BC%80%E6%BA%90%E6%A8%A1%E5%9E%8B%E7%9A%84%E5%AE%89%E5%85%A8%E6%8A%A4%E6%A0%8F%E8%83%BD%E5%8A%9B%E5%88%86%E7%BA%A7red-teaming-%E4%B8%8E%E8%AE%B8%E5%8F%AF%E8%AF%81/</guid><description>开源权重一旦发布便不可撤回，且可被微调绕过对齐。本文讲清能力分级与红队评测方法，给出可运行的护栏代码，并梳理企业采用时的安全与合规清单。</description></item><item><title>Agent 长任务的可靠性：检查点、幂等与断点续跑</title><link>https://zxing2004.github.io/blog/posts/agent-%E9%95%BF%E4%BB%BB%E5%8A%A1%E7%9A%84%E5%8F%AF%E9%9D%A0%E6%80%A7%E6%A3%80%E6%9F%A5%E7%82%B9%E5%B9%82%E7%AD%89%E4%B8%8E%E6%96%AD%E7%82%B9%E7%BB%AD%E8%B7%91/</link><pubDate>Mon, 28 Sep 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/agent-%E9%95%BF%E4%BB%BB%E5%8A%A1%E7%9A%84%E5%8F%AF%E9%9D%A0%E6%80%A7%E6%A3%80%E6%9F%A5%E7%82%B9%E5%B9%82%E7%AD%89%E4%B8%8E%E6%96%AD%E7%82%B9%E7%BB%AD%E8%B7%91/</guid><description>长程异步 Agent 的失败是常态而非例外。本文讲透检查点、幂等与断点续跑三大机制，并给出可直接落地的 Python 实现与工程避坑清单。</description></item><item><title>原生多模态的工程实现：从编码器对齐到统一 token 空间</title><link>https://zxing2004.github.io/blog/posts/%E5%8E%9F%E7%94%9F%E5%A4%9A%E6%A8%A1%E6%80%81%E7%9A%84%E5%B7%A5%E7%A8%8B%E5%AE%9E%E7%8E%B0%E4%BB%8E%E7%BC%96%E7%A0%81%E5%99%A8%E5%AF%B9%E9%BD%90%E5%88%B0%E7%BB%9F%E4%B8%80-token-%E7%A9%BA%E9%97%B4/</link><pubDate>Sat, 19 Sep 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E5%8E%9F%E7%94%9F%E5%A4%9A%E6%A8%A1%E6%80%81%E7%9A%84%E5%B7%A5%E7%A8%8B%E5%AE%9E%E7%8E%B0%E4%BB%8E%E7%BC%96%E7%A0%81%E5%99%A8%E5%AF%B9%E9%BD%90%E5%88%B0%E7%BB%9F%E4%B8%80-token-%E7%A9%BA%E9%97%B4/</guid><description>原生多模态已成旗舰标配，但它与『图像编码器 + 投影层』的拼接方案究竟差在哪？本文拆解 ViT patch 化、模态对齐、视频时序编码等关键实现，并给出可运行的代码骨架与自研/API 决策建议。</description></item><item><title>KV Cache 到底占了多少显存：从 MQA/GQA 到 FP4 量化</title><link>https://zxing2004.github.io/blog/posts/kv-cache-deep-dive/</link><pubDate>Tue, 08 Sep 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/kv-cache-deep-dive/</guid><description>长上下文为何这么吃显存？答案藏在 KV Cache 里。本文从零推导 KV Cache 显存公式，讲清 MHA→MQA→GQA 的演进、FP8/FP4 量化与跨层共享，并给出可复用的显存计算工具。</description></item><item><title>推理服务的成本账本：Prefill/Decode 分离与 Continuous Batching</title><link>https://zxing2004.github.io/blog/posts/inference-cost-batching/</link><pubDate>Tue, 25 Aug 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/inference-cost-batching/</guid><description>Token 经济爆发之后，推理成本成了硬指标。本文从 prefill 与 decode 两种截然不同的瓶颈出发，讲透 Continuous Batching 与 PD 分离为何能提升吞吐，并给出成本核算与指标监控的可运行示例。</description></item><item><title>自托管开源大模型：从权重下载到 vLLM 服务的完整链路</title><link>https://zxing2004.github.io/blog/posts/%E8%87%AA%E6%89%98%E7%AE%A1%E5%BC%80%E6%BA%90%E5%A4%A7%E6%A8%A1%E5%9E%8B%E4%BB%8E%E6%9D%83%E9%87%8D%E4%B8%8B%E8%BD%BD%E5%88%B0-vllm-%E6%9C%8D%E5%8A%A1%E7%9A%84%E5%AE%8C%E6%95%B4%E9%93%BE%E8%B7%AF/</link><pubDate>Fri, 14 Aug 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E8%87%AA%E6%89%98%E7%AE%A1%E5%BC%80%E6%BA%90%E5%A4%A7%E6%A8%A1%E5%9E%8B%E4%BB%8E%E6%9D%83%E9%87%8D%E4%B8%8B%E8%BD%BD%E5%88%B0-vllm-%E6%9C%8D%E5%8A%A1%E7%9A%84%E5%AE%8C%E6%95%B4%E9%93%BE%E8%B7%AF/</guid><description>开源模型密集爆发让自托管成为可选项，但从权重到可服务之间隔着显存规划、并行配置与一堆 OOM 陷阱。本文走通一条从下载、校验到 vLLM 上线与监控的完整链路。</description></item><item><title>万亿参数 MoE 的三个工程支点：专家并行、负载均衡与通信</title><link>https://zxing2004.github.io/blog/posts/%E4%B8%87%E4%BA%BF%E5%8F%82%E6%95%B0-moe-%E7%9A%84%E4%B8%89%E4%B8%AA%E5%B7%A5%E7%A8%8B%E6%94%AF%E7%82%B9%E4%B8%93%E5%AE%B6%E5%B9%B6%E8%A1%8C%E8%B4%9F%E8%BD%BD%E5%9D%87%E8%A1%A1%E4%B8%8E%E9%80%9A%E4%BF%A1/</link><pubDate>Wed, 08 Jul 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E4%B8%87%E4%BA%BF%E5%8F%82%E6%95%B0-moe-%E7%9A%84%E4%B8%89%E4%B8%AA%E5%B7%A5%E7%A8%8B%E6%94%AF%E7%82%B9%E4%B8%93%E5%AE%B6%E5%B9%B6%E8%A1%8C%E8%B4%9F%E8%BD%BD%E5%9D%87%E8%A1%A1%E4%B8%8E%E9%80%9A%E4%BF%A1/</guid><description>以 Kimi K3 全量开源为契机，讲透 MoE 稀疏激活的核心原理，并逐一拆解支撑万亿参数模型落地的三个工程支点：专家并行、负载均衡与通信优化。</description></item><item><title>长文档处理流水线：切片、检索与摘要的工程实践</title><link>https://zxing2004.github.io/blog/posts/%E9%95%BF%E6%96%87%E6%A1%A3%E5%A4%84%E7%90%86%E6%B5%81%E6%B0%B4%E7%BA%BF%E5%88%87%E7%89%87%E6%A3%80%E7%B4%A2%E4%B8%8E%E6%91%98%E8%A6%81%E7%9A%84%E5%B7%A5%E7%A8%8B%E5%AE%9E%E8%B7%B5/</link><pubDate>Mon, 22 Jun 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E9%95%BF%E6%96%87%E6%A1%A3%E5%A4%84%E7%90%86%E6%B5%81%E6%B0%B4%E7%BA%BF%E5%88%87%E7%89%87%E6%A3%80%E7%B4%A2%E4%B8%8E%E6%91%98%E8%A6%81%E7%9A%84%E5%B7%A5%E7%A8%8B%E5%AE%9E%E8%B7%B5/</guid><description>从语义分块到层级摘要再到检索增强，拆解一条可落地的长文档处理流水线，并覆盖表格、代码等特殊内容。</description></item><item><title>拆解 Function Calling：从工具 schema 到多步 Agent 编排</title><link>https://zxing2004.github.io/blog/posts/%E6%8B%86%E8%A7%A3-function-calling%E4%BB%8E%E5%B7%A5%E5%85%B7-schema-%E5%88%B0%E5%A4%9A%E6%AD%A5-agent-%E7%BC%96%E6%8E%92/</link><pubDate>Fri, 15 May 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E6%8B%86%E8%A7%A3-function-calling%E4%BB%8E%E5%B7%A5%E5%85%B7-schema-%E5%88%B0%E5%A4%9A%E6%AD%A5-agent-%E7%BC%96%E6%8E%92/</guid><description>从一个工具定义到多步 Agent 循环，完整拆解 Function Calling 的链路、状态机设计与工程落地要点。</description></item><item><title>百万上下文时代：RAG 还是长上下文？一份工程决策指南</title><link>https://zxing2004.github.io/blog/posts/%E7%99%BE%E4%B8%87%E4%B8%8A%E4%B8%8B%E6%96%87%E6%97%B6%E4%BB%A3rag-%E8%BF%98%E6%98%AF%E9%95%BF%E4%B8%8A%E4%B8%8B%E6%96%87%E4%B8%80%E4%BB%BD%E5%B7%A5%E7%A8%8B%E5%86%B3%E7%AD%96%E6%8C%87%E5%8D%97/</link><pubDate>Fri, 08 May 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/%E7%99%BE%E4%B8%87%E4%B8%8A%E4%B8%8B%E6%96%87%E6%97%B6%E4%BB%A3rag-%E8%BF%98%E6%98%AF%E9%95%BF%E4%B8%8A%E4%B8%8B%E6%96%87%E4%B8%80%E4%BB%BD%E5%B7%A5%E7%A8%8B%E5%86%B3%E7%AD%96%E6%8C%87%E5%8D%97/</guid><description>当百万 token 上下文成为标配，RAG 是否还有必要？本文从原理、成本、准确性到可追溯性给出一套可落地的工程决策框架。</description></item><item><title>RESTful API 设计：容易被忽略的细节</title><link>https://zxing2004.github.io/blog/posts/restful-api-%E8%AE%BE%E8%AE%A1%E5%AE%B9%E6%98%93%E8%A2%AB%E5%BF%BD%E7%95%A5%E7%9A%84%E7%BB%86%E8%8A%82/</link><pubDate>Tue, 10 Mar 2026 00:00:00 +0000</pubDate><author>ricead@qq.com</author><guid>https://zxing2004.github.io/blog/posts/restful-api-%E8%AE%BE%E8%AE%A1%E5%AE%B9%E6%98%93%E8%A2%AB%E5%BF%BD%E7%95%A5%E7%9A%84%E7%BB%86%E8%8A%82/</guid><description>讨论 REST API 设计中常见但容易被忽视的问题，包括资源命名、状态码与版本管理。</description></item></channel></rss>