MindSpeed PTA与MSA环境构建（LLM+MM）

首先需要安装Ascend的hdk驱动，在此不多赘述。

其次需要安装CANN。需要注意的是，CANN 8.5.0正式版的官方教程里声称te、hccl等pip包已经被弃用，有一步是卸载这些包。实测卸载之后会导致mindspore无法正常使用，且CANN官方docker镜像中保留了te等包，因此建议跳过这条卸载命令。

非最新版安装参考（来自sjw学长）：

pta版本MindSpeed-LLM环境安装教程

msa版MindSpeed-LLM安装教程

在此基础之上，可以将其中的版本替换为下方所示：

MindSpore 2.8.0
MindSpeed-Core-MS r0.5.0
Msadapter 0.4.0
CANN 8.5.0
MindSpeed-LLM 2.3.0
MindSpeed-MM 2.3.0
Megatron 12.1

随后对于MindSpeed-MM的安装，参照MindSpeed-MM/docs/zh/pytorch/installation.md · Ascend/MindSpeed-MM – AtomGit | GitCode中的手动安装即可，装入与MindSpeed-LLM的同级目录。

附上“经典款”一键安装脚本（待更新）（前提是安装好内核态驱动），针对LLM的PTA和MSA版：

#!/bin/bash

mkdir /root/do_cyh
cd /root/do_cyh
conda init
source ~/.bashrc

apt update
apt install -y unzip zip htop vim sudo git git-lfs cmake wget curl screen build-essential

mkdir env_tmp
cd env_tmp
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.3.RC1/Ascend-cann-kernels-910b_8.3.RC1_linux-aarch64.deb?response-content-type=application/octet-stream -O Ascend-cann-kernels-910b_8.3.RC1_linux-aarch64.deb
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.3.RC1/Ascend-cann-nnae_8.3.RC1_linux-aarch64.deb?response-content-type=application/octet-stream -O Ascend-cann-nnae_8.3.RC1_linux-aarch64.deb
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.3.RC1/Ascend-cann-nnal_8.3.RC1_linux-aarch64.deb?response-content-type=application/octet-stream -O Ascend-cann-nnal_8.3.RC1_linux-aarch64.deb
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.3.RC1/Ascend-cann-nnrt_8.3.RC1_linux-aarch64.deb?response-content-type=application/octet-stream -O Ascend-cann-nnrt_8.3.RC1_linux-aarch64.deb
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.3.RC1/Ascend-cann-toolkit_8.3.RC1_linux-aarch64.deb?response-content-type=application/octet-stream -O Ascend-cann-toolkit_8.3.RC1_linux-aarch64.deb
echo "Y" | apt install ./Ascend-cann-kernels-910b_8.3.RC1_linux-aarch64.deb
echo "Y" | apt install ./Ascend-cann-nnae_8.3.RC1_linux-aarch64.deb
echo "Y" | apt install ./Ascend-cann-nnal_8.3.RC1_linux-aarch64.deb
echo "Y" | apt install ./Ascend-cann-nnrt_8.3.RC1_linux-aarch64.deb
echo "Y" | apt install ./Ascend-cann-toolkit_8.3.RC1_linux-aarch64.deb
cd ..
rm -rf env_tmp
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source ${HOME}/Ascend/nnal/atb/set_env.sh
npu-smi info

rm -rf /root/miniconda3/envs/pta
rm -rf /root/miniconda3/envs/msa
rm -rf /root/miniconda3/envs/msadapter
rm -rf /root/miniconda3/envs/mindspeed
conda create -n pta python=3.10 -y
conda create -n msa python=3.10 -y
conda init
mkdir pta
cd pta
conda activate pta
pip3 install setuptools==65.7.0
pip3 install protobuf==3.20
wget https://mirrors.nju.edu.cn/pytorch/whl/cpu/torch-2.7.1%2Bcpu-cp310-cp310-manylinux_2_28_aarch64.whl
pip3 install torch-2.7.1+cpu-cp310-cp310-manylinux_2_28_aarch64.whl
wget https://gitcode.com/Ascend/pytorch/releases/download/v7.2.0-pytorch2.7.1/torch_npu-2.7.1-cp310-cp310-manylinux_2_28_aarch64.whl
pip3 install torch_npu-2.7.1-cp310-cp310-manylinux_2_28_aarch64.whl
git clone https://gitcode.com/ascend/MindSpeed.git
cd MindSpeed
git checkout 2.2.0_core_r0.12.1
pip install -r requirements.txt
pip3 install -e .
cd /root/do_cyh/pta
git clone https://gitcode.com/ascend/MindSpeed-LLM.git 
cd MindSpeed-LLM
git checkout 2.2.0
cd /root/do_cyh/pta
git clone https://gitee.com/mirrors/Megatron-LM.git
cd Megatron-LM
git checkout core_v0.12.1
cp -r megatron ../MindSpeed-LLM/
cd /root/do_cyh/pta
cd MindSpeed-LLM
pip3 install -r requirements.txt
pip3 install -e .
cd /root/do_cyh/pta
git clone -b master https://gitcode.com/Ascend/apex.git
cd apex
bash scripts/build.sh --python=3.10
pip3 uninstall apex
pip3 install --upgrade ./apex/dist/apex-0.1+ascend-cp310-cp310-linux_aarch64.whl 
cd ../..
export PYTHONPATH=/root/do_cyh/pta/Megatron-LM:$PYTHONPATH
pip install ruamel.yaml colorama

cd /root/do_cyh
mkdir msa
cd msa
conda activate msa
pip3 install setuptools==65.7.0
wget https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.7.1/MindSpore/unified/aarch64/mindspore-2.7.1-cp310-cp310-linux_aarch64.whl -O mindspore-2.7.1-cp310-cp310-linux_aarch64.whl
pip install ./mindspore-2.7.1-cp310-cp310-linux_aarch64.whl
git clone https://gitcode.com/ascend/MindSpeed-Core-MS.git -b r0.4.0
cd MindSpeed-Core-MS
pip install -r requirements.txt
pip install colorama
source auto_convert.sh llm
cd ..

conda activate base
pip install colorama
conda create -n mindspeed --clone /root/miniconda3/envs/pta
conda create -n msadapter --clone /root/miniconda3/envs/msa

未完待续。。。。。。。