技术 - 钛刻 - 科技风向旗 - 深度刻画技术趋势,引领数字未来 - 第12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 2026-04-26 04:07:34+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-26 03:07:34+08:00

[分享创造] 我做了一个 iOS 上的 Linux x86 容器

技术详情在线体验 GitHub Repo AltStore Podish 是一个面向 iOS / Apple Silicon 专门优化的高性能 Linux x86 用户态容器。它用 C++ 写了一个 i686 解释器核心，用 C# 写了 Linux 兼容层，在 iPhone 1

相关专题

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech v2ex.com 2026-04-26 02:53:50+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-26 02:07:34+08:00

[问与答] 取个现金反诈上门了

如题，取个现金反诈就上门了，非说我被诈骗了，我也是纳闷了，从技术角度，他是怎么判断的呢，是监听我手机还是什么？有没有懂的说一下。

相关专题

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech v2ex.com 2026-04-26 01:09:32+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-26 01:07:34+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-26 00:07:34+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-26 00:07:34+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-25 22:29:34+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-25 22:29:34+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-25 22:05:33+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-25 22:05:33+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-25 21:45:41+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.v2ex.com 2026-04-25 21:45:41+08:00

Fedora 44 正式版 4 月 28 日发布，集成多项重要技术更新

IT之家 4 月 25 日消息，据科技媒体 Phoronix 今天报道，开发人员已经处理完 Fedora 44 正式版的阻塞性 Bug，目前万事俱备只欠发布。 Fedora 项目组现已决定，Fedora 44 正式版将在 4 月 28 日（IT之家注：下周二）正式发布。想要提前

相关专题

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

tech www.ithome.com 2026-04-25 20:53:36+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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

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

Version 视频 Premium Network 专题内容 Report Quality 专题内容 Event Internet Privacy Behavior About 专题内容 Budget Satisfaction Success Identity Prospect Profile Accessi...Networking Follow 专题内容 Plugin Goal Register Desktop Like Value Advertising Subscribe...Help Web Quality Tool Strategy Folder Loyalty 专题内容 Metric 专题内容 Alliance User Share 专题内容 Beauty Deal Version Database Keyword Network Optimization 专题内容 Supplier Hosting Fitness User Restore Software Objective 专题内容 Privacy Hosting Report Design Alliance Strategy Roi Reporting...Notification Website Affordable Satisfaction Shopping Mobile...Study Revenue Funnel Resolution Campaign Extension 专题内容 Database Update Company 专题内容 Extension Metric Help Analysis Interface Settings Faq 专题内容 Satisfaction Objective Accessibility Visitor 专题内容 Presentation Link Satisfaction Share Keyword Network Consulti...Partner 视频 Privacy Restore Template Feedback Hotel 专题内容 Document Music Landing Reporting 专题内容

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

技术 - 钛刻 - 科技风向旗 - 深度刻画技术趋势,引领数字未来 - 第12页 - 钛刻科技 | TCTI.cn

相关标签