Add MACA
parent
727a366b70
commit
d7970e6950
298
MACA.md
Normal file
298
MACA.md
Normal file
@ -0,0 +1,298 @@
|
||||
# MACA SDK
|
||||
|
||||
### 适配 Centos、Kylin 系统
|
||||
|
||||
- **配置YUM源**
|
||||
|
||||
```bash
|
||||
cat <<EOF> /etc/yum.repos.d/maca-sdk-rpm.repo
|
||||
[maca-sdk]
|
||||
name=Maca Sdk Yum Repository
|
||||
baseurl=https://repos.metax-tech.com/r/maca-sdk-rpm-$(uname -m)/
|
||||
enabled=1
|
||||
gpgcheck=0
|
||||
EOF
|
||||
|
||||
yum makecache
|
||||
```
|
||||
|
||||
#### 安装MACA SDK
|
||||
|
||||
- **安装maca_sdk最新版本**
|
||||
```bash
|
||||
yum install maca_sdk
|
||||
```
|
||||
|
||||
- **安装maca_sdk指定版本**
|
||||
```bash
|
||||
yum --showduplicates list |grep maca_sdk
|
||||
yum install maca_sdk-<version>-<release>
|
||||
```
|
||||
|
||||
- **升级maca_sdk到最新版本**
|
||||
```bash
|
||||
yum update maca_sdk
|
||||
```
|
||||
|
||||
- **卸载maca_sdk**
|
||||
```bash
|
||||
yum autoremove maca_sdk
|
||||
```
|
||||
|
||||
|
||||
|
||||
### 适配 Ubuntu 系统
|
||||
|
||||
- **配置APT源**
|
||||
```bash
|
||||
curl -fsSL https://repos.metax-tech.com/public.gpg.key | apt-key add -
|
||||
echo "deb [arch=$(dpkg --print-architecture)] https://repos.metax-tech.com/r/maca-sdk-deb/ stable main" | tee /etc/apt/sources.list.d/maca-sdk-deb.list
|
||||
apt-get update
|
||||
|
||||
```
|
||||
|
||||
#### 安装MACA SDK
|
||||
|
||||
- **安装maca_sdk最新版本**
|
||||
```bash
|
||||
apt install maca_sdk
|
||||
```
|
||||
|
||||
- **安装maca_sdk指定版本**
|
||||
```bash
|
||||
apt list --all-versions |grep maca_sdk
|
||||
apt install maca_sdk=<version>
|
||||
```
|
||||
|
||||
- **升级maca_sdk到最新版本**
|
||||
```bash
|
||||
apt install --only-upgrade maca_sdk
|
||||
```
|
||||
|
||||
- **卸载maca_sdk**
|
||||
```bash
|
||||
apt autoremove maca_sdk
|
||||
```
|
||||
|
||||
|
||||
# Metax Driver
|
||||
|
||||
|
||||
### 平台兼容列表
|
||||
|
||||
| CPU架构 | 操作系统 | 内核版本 | 状态 |
|
||||
|-----------|----------------|------------------------------------|----------------------|
|
||||
| x86_64 | Ubuntu 18.04 | 5.4.0-42-generic | 支持 |
|
||||
| | | 5.4.0-131-generic | |
|
||||
| | | 5.4.0-148-generic | |
|
||||
| x86_64 | Ubuntu 20.04 | 5.4.0-26-generic | 支持 |
|
||||
| | | 5.4.0-42-generic | |
|
||||
| | | 5.15.0-58-generic | |
|
||||
| x86_64 | Ubuntu 22.04 | 5.15.0-72-generic | 支持 |
|
||||
| | | 5.19.0-46-generic | |
|
||||
| x86_64 | CentOS 8 | 4.18.0-240.el8.x86_64 | 支持 |
|
||||
| x86_64 | RedHat 8.3 | 4.18.0-240.el8.x86_64 | 支持 |
|
||||
| x86_64 | RedHat 8.6 | 4.18.0-372.9.1.el8.x86_64 | 支持 |
|
||||
| x86_64 | RedHat 9 | 5.14.0-70.22.1.el9_0.x86_64 | 支持 |
|
||||
| x86_64 | CentOS 7 | 4.19.0-1.el7.elrepo.x86_64 | 支持加载Docker Container形式 |
|
||||
| | | 5.14.0-7.x86_64 | |
|
||||
| x86_64 | BC Linux | 4.19.90-2107.6.0.0100.oe1.bclinux | 支持 |
|
||||
| x86_64 | BCLinux R8 U2 | 4.19.0-240.23.11.el8_2.bclinux.x86_64 | 支持 |
|
||||
| x86_64 | CC Linux | 5.15.131-2.cl9.x86_64 | 支持 |
|
||||
| x86_64 | Kylin V10 SP2 | 4.19.90-24.4.v2101.ky10.x86_64 | 支持 |
|
||||
| x86_64 | ALinux3 | 5.10.134-13.1.al8.x86_64 | 支持 |
|
||||
| x86_64 | CTYun 23.01 | 5.10.0-136.12.0.86.ctl3.x86_64 | 支持 |
|
||||
| x86_64 | KeyarchOS 5.8 | 4.19.91-27.4.19.kos5.x86_64 | 支持 |
|
||||
|
||||
|
||||
### 系统兼容性要求
|
||||
|
||||
- 若曦云系列GPU无法识别为PCIe设备(可通过 `lspci | grep 9999` 进行查看),需关闭BIOS里兼容性支持模块 (CSM)选项。
|
||||
- 若PCIe BAR需要支持4GB以上地址空间,需打开BIOS里Large Bar选项。
|
||||
- 每张曦云系列GPU板卡需要三个BAR,空间大小分别为1 MB、8 MB和64 GB。
|
||||
|
||||
|
||||
### 环境检查(以Ubuntu 18.04为例)
|
||||
|
||||
| 序号 | 检查项目 | 检查命令 | 说明 |
|
||||
|------|---------------------|-------------------------------|---------------------------------------------------------------------------------------------------------------|
|
||||
| 1 | 服务器CPU架构、操作系统、内核版本 | uname -m, lsb_release -a, uname -r | 确认CPU架构、操作系统和内核版本是否在平台兼容列表中, 不匹配则需更新环境 |
|
||||
| 2 | 系统是否安装过驱动 | lsmod \| grep metax | 若无内容显示, 表示未安装软件包, 可直接安装<br/> 若有内容显示, 表示已安装, 需要卸载后再安装 |
|
||||
| 3 | 板卡是否在位 | lspci \| grep 9999 | 确认GPU板卡正常在位, 含“9999”的行数应与板卡数量匹配 |
|
||||
| 4 | udev 配置 | cat /etc/udev/rules.d/70-metax.rules | 若有群组管理需要: <br/> 只允许 video 组的成员使用曦云 GPU, 则配置文件有如下内容: <br/> KERNEL=="mxcd", GROUP="video", MODE="0660" <br/> KERNEL=="renderD[0-9]*", GROUP="video", MODE="0660" <br/> 若没有以上内容, 则表示只有 root 用户可以使用曦云 GPU。<br/> 若无群组管理需要, 可删除此配置文件 (deb 包安装会自动创建此配置文件) |
|
||||
| 5 | 虚拟化 | dmesg | grep "no space" | 确认系统是否支持 SRIOV 虚拟化 |
|
||||
| 6 | IOMMU配置 | vim /etc/default/grub | 确保 IOMMU 配置生效, 添加 iommu.passthrough=1, 更新 grub 并重启系统 |
|
||||
| 7 | gpu_sched模块 | modinfo gpu_sched | 确认是否缺少依赖模块, 若缺少则安装内核可选模块包 |
|
||||
| 8 | 第三方驱动加载 | modprobe metax | 在 SLES 系统中, 确认是否允许第三方驱动加载 |
|
||||
|
||||
|
||||
|
||||
|
||||
### 配置YUM源
|
||||
|
||||
- **适配 Centos7 系统**
|
||||
|
||||
```bash
|
||||
cat <<EOF> /etc/yum.repos.d/metax-driver-centos7.repo
|
||||
[metax-centos7]
|
||||
name=Maca Driver Yum Repository
|
||||
baseurl=https://repos.metax-tech.com/r/metax-driver-centos7-$(uname -m)/
|
||||
enabled=1
|
||||
gpgcheck=0
|
||||
EOF
|
||||
|
||||
yum makecache
|
||||
```
|
||||
|
||||
- **适配 CentOS 8、CentOS 9 及 Kylin 系统**
|
||||
```bash
|
||||
cat <<EOF> /etc/yum.repos.d/metax-driver-centos.repo
|
||||
[metax-centos]
|
||||
name=Maca Driver Yum Repository
|
||||
baseurl=https://repos.metax-tech.com/r/metax-driver-centos-$(uname -m)/
|
||||
enabled=1
|
||||
gpgcheck=0
|
||||
EOF
|
||||
|
||||
yum makecache
|
||||
```
|
||||
|
||||
##### YUM安装metax-driver
|
||||
|
||||
- **安装metax-driver驱动**
|
||||
```bash
|
||||
yum install metax-driver
|
||||
```
|
||||
|
||||
- **若需要启用SRIOV,需要安装额外的虚拟化支持包 mxgvm,以启用 SR-IOV 功能**
|
||||
```bash
|
||||
yum install metax-driver mxgvm
|
||||
```
|
||||
- **卸载metax-driver**
|
||||
```bash
|
||||
yum autoremove metax-driver
|
||||
```
|
||||
|
||||
### 配置APT源
|
||||
|
||||
- **适配所有 Ubuntu 发行版本系统**
|
||||
```bash
|
||||
curl -fsSL https://repos.metax-tech.com/public.gpg.key | apt-key add -
|
||||
echo "deb [arch=$(dpkg --print-architecture)] https://repos.metax-tech.com/r/metax-driver-ubuntu/ stable main" | tee /etc/apt/sources.list.d/metax-driver-ubuntu.list
|
||||
apt-get update
|
||||
|
||||
##### APT安装metax-driver
|
||||
|
||||
- **安装metax-driver驱动**
|
||||
```bash
|
||||
apt install metax-driver
|
||||
```
|
||||
|
||||
- **若需要启用SRIOV,需要安装额外的虚拟化支持包 mxgvm,以启用 SR-IOV 功能**
|
||||
```bash
|
||||
apt install metax-driver mxgvm
|
||||
```
|
||||
|
||||
- **卸载metax-driver**
|
||||
```bash
|
||||
apt autoremove --purge metax-driver
|
||||
```
|
||||
|
||||
|
||||
|
||||
|
||||
### 验证安装
|
||||
- **在安装完成后需要重启服务器**
|
||||
```bash
|
||||
reboot
|
||||
```
|
||||
|
||||
- **重启服务器后,可以通过以下命令确认驱动是否成功加载,如果看到相关驱动模块列出,说明安装已成功**
|
||||
```bash
|
||||
lsmod | egrep "(metax|mxgvm)"
|
||||
```
|
||||
|
||||
- **安装 metax-driver,安装过程中将包含四个软件包：**
|
||||
- **metax-driver**
|
||||
- **metax-linux**
|
||||
- **mxsmt**
|
||||
- **mxfw**
|
||||
|
||||
- **检查已安装包,使用以下命令检查相关包的安装情况：**
|
||||
|
||||
```bash
|
||||
dpkg -l | egrep "(metax|mxsmt|mxfw)"
|
||||
rpm -qa | egrep "(metax|mxsmt|mxfw)"
|
||||
```
|
||||
|
||||
- **Driver 软件包内容清单**
|
||||
| 文件名 | 说明 |
|
||||
|------------------------------------|----------------------------------------------|
|
||||
| metax-linux_x.x.x-xxx_amd64.deb | 曦云系列 GPU KMD 驱动、工具及相关配置文件 |
|
||||
| mxgvm_x.x.x-xxx_amd64.deb | 曦云系列 GPU Virtualization Manager、工具及相关配置文件 |
|
||||
| mxfw_x.x.x.x.all.deb | 曦云系列 GPU 固件包 |
|
||||
| mxsmt_x.x.x.x.amd64.deb | mx-smi 系统管理工具,MXSML 系统管理库 |
|
||||
|
||||
|
||||
|
||||
# MACA Pytorch
|
||||
|
||||
### 环境准备
|
||||
|
||||
#### Python 环境
|
||||
- 匹配目标安装包的 Python 版本(Python 3.8 、Python 3.10、Python 3.12)
|
||||
|
||||
#### MXMACA 环境
|
||||
- 安装 Driver 软件包以及 MXMACA SDK 软件包
|
||||
|
||||
#### cu-bridge 环境
|
||||
- 在安装 MXMACA 环境到 `/opt/maca/` 目录后，将 cu-bridge 解压到 `/opt/maca/tools/` 目录下，具体操作步骤参见相关文档。可以检验 `/opt/maca/tools/cu-bridge/bin/cucc` 文件是否存在，从而判断 cu-bridge 环境是否配置完成。
|
||||
- cu-bridge 环境不需要安装 CUDA Toolkit。
|
||||
- 如果不使用 PyTorch Extension 功能，可以不配置 cu-bridge 环境。关于 PyTorch Extension 功能，参见官网描述。
|
||||
|
||||
|
||||
### Pip安装
|
||||
#### 单包安装
|
||||
- **使用以下命令安装指定包**
|
||||
```bash
|
||||
pip install <package> -i https://repos.metax-tech.com/r/maca-pypi/simple --trusted-host repos.metax-tech.com
|
||||
```
|
||||
|
||||
#### 使用requirements.txt批量安装
|
||||
- **在命令行中指定 maca-pypi 源地址**
|
||||
```bash
|
||||
pip install -r ./requirements.txt -i https://repos.metax-tech.com/r/maca-pypi/simple --trusted-host repos.metax-tech.com
|
||||
```
|
||||
```bash
|
||||
cat <<EOF> ./requirements.txt
|
||||
apex
|
||||
dropout_layer_norm
|
||||
flash_attn
|
||||
fused_dense_lib
|
||||
rotary_emb
|
||||
torch
|
||||
torchaudio
|
||||
torchvision
|
||||
triton
|
||||
xentropy_cuda_lib
|
||||
xformers
|
||||
|
||||
EOF
|
||||
```
|
||||
|
||||
|
||||
#### 列出可用版本
|
||||
|
||||
- **对于 pip 22.0 及之前版本,使用以下命令触发错误并列出包的所有可用版本**
|
||||
```bash
|
||||
pip install <package>== -i https://repos.metax-tech.com/r/maca-pypi/simple --trusted-host repos.metax-tech.com
|
||||
```
|
||||
|
||||
- **对于 pip 22.0 之后版本,使用以下命令列出某个包的所有可用版本**
|
||||
```bash
|
||||
pip index versions <package> -i https://repos.metax-tech.com/r/maca-pypi/simple --trusted-host repos.metax-tech.com
|
||||
```
|
||||
|
||||
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user