Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Deploying Qwen3.5 4B on a Legacy RX 580 Server

AI summary

A report from 掘金 · 人工智能 documents running Qwen3.5 4B with Ollama 0.40.14 on Windows 10 using an Intel Xeon E5-2670 v2, 32 GB of memory, and an 8 GB AMD RX 580. The author exposes the model to WorkBuddy and Dify through an OpenAI-compatible API, with the Xeon serving as fallback compute at a reported prefill speed of about 54–74 tokens per second. Because the Vulkan flash-attention path produced zero-filled output on longer prompts, the deployment disables flash attention and accepts slower generation for more predictable results.

In-depth insight

The first-hand configuration documents a reproducible way to serve Qwen3.5 4B on aging, low-resource hardware, including its API setup, context trade-offs, concurrency limits, and troubleshooting steps.

先说结论:一台十年前的机器,能跑、能对外提供服务,但整个过程不是"装完就完事"——这代硬件(AMD Polaris 显卡 + 没有 AVX2 的 Ivy Bridge CPU)逼出了三个反直觉的决定:关掉 flash attention、接受 CPU 当主力、用固定别名把参数钉死。

环境:Intel Xeon E5-2670 v2 / 32 GB DDR3L / AMD RX 580 2048SP 8 GB / Windows 10 专业版 19045,跑 Ollama 0.40.14 + 千问 3.5 4B(Q4_K_M),给局域网内的 WorkBuddy、Dify 等平台提供 OpenAI 兼容 API。

一、硬件直接决定了方案边界

项目实际配置对方案的影响
CPUXeon E5-2670 v2,10 核 20 线程,Ivy Bridge,仅 AVX 无 AVX2只能当兜底算力,预填充约 54~74 tok/s
内存32 GB DDR3L-1600撑得住 9.6 GB 的模型 + KV 占用
显卡RX 580 2048SP / 8 GB,Polaris gfx803ROCm 不支持(只支持 RX 7600 起的 RDNA3),只能走 Vulkan
系统Windows 10 19045Ollama 0.40.x 默认启用 Vulkan 后端,Polaris 可用

#Qwen3.5 4B#Ollama#WorkBuddy#Dify#AMD RX 580#Intel Xeon E5-2670 v2#ROCm#Vulkan

Original · 8 min readRead the original
Group chat

No comments yet. Start the conversation.