如何安装Tensorflow和GPU配置

课题组某一台服务器升级后，很多环境丢失了，4块3090的GPU的驱动已安装好，但没有公用的Tensorflow可使用。于是自己鼓捣了一番Tensorflow的安装，等管理员安装公用的环境不知道要到猴年马月……

服务器是Linux系统（CentOS），GPU是英伟达公司的3090，已经安装好驱动，可以通过命令看到相关信息：

$ nvidia-smi
Tue May 28 20:54:09 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.78                 Driver Version: 550.78         CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3090        On  |   00000000:25:00.0 Off |                  N/A |
| 31%   30C    P8             18W /  350W |       2MiB /  24576MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

目标是：安装Tensorflow 2 并调用GPU加速。

一、确定要安装的各软件的版本

要安装Tensorflow并调用GPU，需要安装：tensorflow2(2版本的gpu和cpu版本合并在一个安装包里，1版本需要分别安装gpu和cpu版本，这里不做讨论），cudnn和cudatoolkit。此外还需要安装高能物理领域常用的ROOT软件。

cudatoolkit：The toolkit includes GPU-accelerated libraries, debugging and optimization tools, a C/C++ compiler, and a runtime library. 安装的版本低于GPU信息里的cuda版本（上方代码块中显示为12.4）。

cudnn：选择与cudatoolkit兼容的版本，可查阅nvidia官网：https://developer.nvidia.com/rdp/cudnn-archive

python、tensorflow的版本：可根据https://www.tensorflow.org/install/source?hl=zh-cn#gpu选择。

这里我们选择cudatoolkit 11.2，cudnn 8.1（满足cuda 11.x，从cuda-cudnn GPU版本对照表里挑了一个cuda11.2对应的cudnn版本8.1，配置GPU应该没问题），python 3.8（比较常见的版本），tensorflow 2.9（之前服务器上的版本），root 6.26（常见的root版本）。gcc版本似乎在7以上就行（包管理器基于我以上环境安装了10.3）。以上版本在conda channel里都能search到，比较方便。

此外还需要配置相关环境变量。下面会讨论环境变量的设置。

二、安装

安装基于conda虚拟环境，能避免不同环境下的依赖冲突。

1. 新建一个conda环境tf_env，并安装python3.8：

   conda create -n tf_env python=3.8 -c conda-forge

2.  搜索和安装cudatoolkit 11.2：

   conda search cudatoolkit=11.2
   conda install cudatoolkit=11.2

3. 安装cudnn 8.1:

   conda install cudnn=8.1

4. 安装python 3.8：

   conda install python=3.8

5. 安装tensorflow2.9，这里用官方推荐的pip安装，和conda类似。 我这里需要指定URL pypi.org/simple，否则解析错误。pip会自动安装很多tensorflow依赖的包，可以通过conda list查看该环境下已安装的包。

   pip install tensorflow==2.9 -i https://pypi.org/simple

6. 配置环境变量。tensorflow已安装完成。如果成功读取gpu且返回张量，则成功：

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/
   python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
   # 
   2024-05-28 22:08:21.114048: I tensorflow/core/util/util.cc:169] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
   2024-05-28 22:08:24.152377: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F AVX512_VNNI FMA
   To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
   2024-05-28 22:08:28.675827: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:42] Overriding orig_value setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.
   2024-05-28 22:08:28.675894: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 22301 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 3090, pci bus id: 0000:25:00.0, compute capability: 8.6
   2024-05-28 22:08:28.689181: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:42] Overriding orig_value setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.
   2024-05-28 22:08:28.689211: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:1 with 22301 MB memory:  -> device: 1, name: NVIDIA GeForce RTX 3090, pci bus id: 0000:5b:00.0, compute capability: 8.6
   2024-05-28 22:08:28.690197: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:42] Overriding orig_value setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.
   2024-05-28 22:08:28.690226: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:2 with 22301 MB memory:  -> device: 2, name: NVIDIA GeForce RTX 3090, pci bus id: 0000:9b:00.0, compute capability: 8.6
   2024-05-28 22:08:28.690678: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:42] Overriding orig_value setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.
   2024-05-28 22:08:28.690713: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:3 with 22301 MB memory:  -> device: 3, name: NVIDIA GeForce RTX 3090, pci bus id: 0000:c8:00.0, compute capability: 8.6
   tf.Tensor(2062.545, shape=(), dtype=float32)

7. 安装root6.26， 这一步推荐用mamba管理器安装，预先编译成c++，安装更快。否则conda安装root解析环境巨慢。这一步会自动安装root依赖的gcc（默认安装了10.3）和很多其他库。

   conda install mamba
   mamba install root=6.26.6

   $ conda list | grep "gcc"
   _libgcc_mutex             0.1                 conda_forge    conda-forge
   gcc                       10.3.0              he2824d0_10    https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
   gcc_impl_linux-64         10.3.0              hf2f2afa_16    https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
   gcc_linux-64              10.3.0              hc39de41_10    https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
   libgcc-devel_linux-64     10.3.0              he6cfe16_16    https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
   libgcc-ng                 13.2.0               h77fa898_7    conda-forge

   至此，ROOT安装完成，可在python中import ROOT测试：

8. >>> import ROOT as rt
   >>> rt.__version__
   '6.26/08'