GB - 钛刻 - 科技风向旗 - 深度刻画技术趋势,引领数字未来 - 第8页 - 钛刻科技

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

tech www.v2ex.com 2026-04-25 20:44:05+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 20:44:05+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 20:44:05+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 20:44:05+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 20:20:13+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 20:06:01+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 19:04:06+08:00

[推广] [分享] 推荐一个极高性价比的海外代理 Novproxy：$0.5/GB 起， 1 亿+ 住宅 IP 资源

各位 V 友，做跨境电商、海外社媒矩阵或者大规模爬虫任务的朋友，对代理 IP 的消耗肯定深有体会。大厂虽然稳，但流量费和起步门槛确实高。最近在用 Novproxy ，感觉整体性价比在同类产品中非常有竞争力，特别适合对成本控制有要求的开发者和工作室。分享几个我比较看中的点： 💎

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech v2ex.com 2026-04-25 17:42:48+08:00

苹果+Visa用户可以关注一下：我们能从锁着的苹果手机里偷走一万刀吗？

【我们能从锁着的苹果手机里偷走一万刀吗？】 https://www.bilibili.com/video/BV1sRogB1EVy 看了一下还是有一点点吓人的虽然热评说不用担心 1 个帖子 - 1 位参与者阅读完整话题

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech linux.do 2026-04-25 17:22:14+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 17:04:33+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 17:04:33+08:00

英伟达已适配 DeepSeek-V4 AI 模型，GB200 NVL72 开箱性能超 150 tokens / sec / user

IT之家 4 月 25 日消息，英伟达今天（4 月 25 日）发布博文，宣布其 NVIDIA Blackwell 平台已适配 DeepSeek-V4-Pro 与 DeepSeek-V4-Flash 两款模型，开发者可通过 NVIDIA NIM 微服务下载部署，或利用 SGLa

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.ithome.com 2026-04-25 15:36:02+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 15:17:18+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 15:17:18+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 14:43:04+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 13:56:57+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 13:56:57+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 13:36:22+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 13:03:16+08:00

我做了个工具让 8GB 显卡跑 30B 模型从 3 tok/s 提到 21 tok/s，记录一下技术发现

最近在折腾本地大模型，发现一个核心问题：Ollama 和 LM Studio 能让模型跑起来，但参数全靠猜——上下文长度、KV cache 类型、MoE expert 放哪、ubatch 多大……用默认参数基本是在浪费显卡。于是做了个工具自动找最优配置，过程中踩了不少坑，记录一

相关专题

Faq Careers Company Conversion Plugin Collaboration Meeting O...Management Quality Section Advertising 专题内容 Training Support Security Machine 专题内容 Button 视频 App 专题内容 Contact Roi 专题内容 Policy Hotel Course Tutorial 专题内容 URL 专题内容 Demographic 专题内容 Browser Form Supplier Guide 游戏 Value Hosting Objective Unsubs...Demographic Global Restaurant 专题内容 Engagement Networking Expensive Software System Income Revenu...Analytics Integration Whitepaper Help Technology Innovation T...Networking Course Report Premium Notification Demographic 专题内容 Satisfaction Growth Label Global Tool Plugin About Music Cons...Company Roi Community Webinar Email Optimization 专题内容 Plugin Presentation Guide Development Roi Prospect Like 专题内容 Project Upload Audience Lead Email Cloud Screen Platform Care...Dashboard 专题内容 User Community 专题内容 Networking Comment Digital Calculator Tutorial Module Deadlin...

tech www.v2ex.com 2026-04-25 12:52:52+08:00

GB - 钛刻 - 科技风向旗 - 深度刻画技术趋势,引领数字未来 - 第8页 - 钛刻科技 | TCTI.cn

相关标签