From 5ee45ecdda306fc720fc9f69efaaa5b5cdbfda23 Mon Sep 17 00:00:00 2001 From: Dipam Paul Date: Thu, 7 Aug 2025 12:11:46 -0700 Subject: [PATCH] docs(Atom-7B-chat-WebDemo): add flash_attn installation note with wheel link for PyTorch 2.0 + CUDA 11.8 (fixes #440) --- models/Atom/01-Atom-7B-chat-WebDemo.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/models/Atom/01-Atom-7B-chat-WebDemo.md b/models/Atom/01-Atom-7B-chat-WebDemo.md index 7af7b4e..99620eb 100644 --- a/models/Atom/01-Atom-7B-chat-WebDemo.md +++ b/models/Atom/01-Atom-7B-chat-WebDemo.md @@ -17,6 +17,15 @@ python -m pip install --upgrade pip pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple #安装项目相关依赖包 pip install modelscope==1.9.5 transformers==4.35.2 gradio==4.4.1 SentencePiece==0.1.99 accelerate==0.24.1 bitsandbytes==0.41.2.post2 + +# 如遇 flash_attn 安装失败，请根据环境手动安装匹配的 wheel 包 +# 例如：PyTorch 2.0.0 + Python 3.8 + CUDA 11.8（Ubuntu 20.04） +# 可直接使用下述预编译 wheel： +# 参考来源：https://github.com/Dao-AILab/flash-attention/releases +pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu118torch2.0cxx11abiFALSE-cp38-cp38-linux_x86_64.whl + +# 其他环境请从 flash-attention Releases 页面选择与自身 CUDA / PyTorch / Python 版本匹配的 wheel +# 若无匹配 wheel，亦可参考官方文档源码编译安装 ``` ## 模型下载