windows官方整合包5090等显卡小白提速教程

直接运行整合包，然后升级CUDA12.8可以成功，也能使用，但是会有一堆错误，而且Xformers，Flash Attn，Sage Attn也不会调用。

所以有个最简单的办法。

首先你找个别人发布在windows中可以在5090上使用Xformers的COMFYUI的python包，版本可以是3.12或别的都行，bilibili上就能搜到。

1.把\system\python的python全部删除，换成5090可以用的COMFYUI的python包

2.在新的\system\python目录用输入CMD打开终端

3.依次逐条输入下面升级命令

python -m pip install --upgrade diffusers

python -m pip install --upgrade accelerate

python -m pip install --upgrade transformers

python -m pip install --upgrade gradio

python -m pip install --upgrade numpy

python -m pip install --upgrade scipy

python -m pip install --upgrade sentencepiece

python -m pip install --upgrade pillow

python -m pip install --upgrade requests

python -m pip install --upgrade torchsde

4.安装Flash Attn

首先安装ninja，否者你在windows下面编译Flash Attn可能要10多个小时

python -m pip install ninja

最后安装Flash Attn

python -m pip install flash-attn --no-build-isolation

编译速度区别于你的PC配置，几分钟到2小时

如果你发现系统只有2个进程在编译，把你的python和python\Scripts目录添加进系统环境变量

然后用MAX_JOBS=10 python -m pip install flash-attn --no-build-isolation进行编译，速度就会很快MAX_JOBS=多少取决于你有几个CPU核心

5.回到整合包根目录

运行run.bat

如果你看到

Xformers is installed!
Flash Attn is installed!
Sage Attn is installed!

然后下面也没提示错误，说明你的5090等显卡能全速运行

5090等显卡运行效率会是目前在COMFYUI上工作流的数倍

Review the issue body as the proposed Windows 5090 setup procedure; no repository documentation file is named. Check how the bundled system\python environment and run.bat are documented, then capture the dependency installation steps and the expected Xformers, Flash Attn, and Sage Attn messages as the completion criteria.

Written by the indexing model from the issue text.