ubuntu16.04一步到位安装CUDA10.02、PyTorch、PaddlePaddle、TensorRT

由于很多框架和套件停止了10.0的支持，将原始已安装完成的cuda10.0的服务器重新配置cuda10.2和新的驱动

原始驱动卸载

1、停X Server

sudo service lightdm stop

2、卸载之前的Driver

sudo /usr/bin/nvidia-uninstall

卸载完成后输入nvidia-smi应提示找不到命令

nvidia-smi: command not found

CUDA安装

点击此处进入CUDA下载页面

选择需要的版本

根据安装提示，输入如下命令

wget https://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
sudo sh cuda_10.2.89_440.33.01_linux.run

输入accept

┌──────────────────────────────────────────────────────────────────────────────┐
│  End User License Agreement                                                  │
│  --------------------------                                                  │
│                                                                              │
│                                                                              │
│  Preface                                                                     │
│  -------                                                                     │
│                                                                              │
│  The Software License Agreement in Chapter 1 and the Supplement              │
│  in Chapter 2 contain license terms and conditions that govern               │
│  the use of NVIDIA software. By accepting this agreement, you                │
│  agree to comply with all the terms and conditions applicable                │
│  to the product(s) included herein.                                          │
│                                                                              │
│                                                                              │
│  NVIDIA Driver                                                               │
│                                                                              │
│                                                                              │
│  Description                                                                 │
│                                                                              │
│  This package contains the operating system driver and                       │
│──────────────────────────────────────────────────────────────────────────────│
│ Do you accept the above EULA? (accept/decline/quit):                         │
│ accept                                                                       │
└──────────────────────────────────────────────────────────────────────────────┘

我们驱动和cuda分开装，安装时通过Enter取消driver的勾选，如下：

安装完成后，修改环境变量并激活，如下

vim ~/.bashrc

export PATH="/usr/local/cuda-10.2/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda-10.2/lib64:$LD_LIBRARY_PATH"
export CUDA_HOME=usr/local/cuda-10.2$CUDA_HOME

source ~/.bashrc

输入nvcc -V验证

nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89

出现10.2表示安装成功！！！

CUDNN配置

下载适配CUDA版本的cudnn，我这里下载8.2的版本，为了与后面的TensorRT匹配。点击此处进入官方下载连链接，选择Library包

右键复制链接

下载完后使用mv命令重命名（去掉tgz后面的字段）

mv cudnn-10.2-linux-x64-v8.2.4.15.tgz\?QCogpoDS_Z1q53hHFgWmiQWSZqnkTJQ7rwBVj-b9kZc0nkobdCw11ri1t9znzIPSTKUiijoxTwJBGOCxYUj_3QHjlTERLhG8lOFu63haEzcKbGjnUAQDxG6SiJJvoySYHG-8Su7EpOGKoo5iJySeHSnvdJrXp2BhnIKmYKqokryTSw9WVWL5lkZ9AeoSuRr4nPltqp_jV  cudnn-10.2-linux-x64-v8.2.4.15.tgz

如此，你应该具有如下几个文件

解压缩，解压后出现cuda文件夹

tar zxvf cudnn-10.2-linux-x64-v8.2.4.15.tgz

复制相应的文件到之前安装的CUDA目录

sudo cp cuda/include/* /usr/local/cuda-10.2/include
sudo cp cuda/lib64/* /usr/local/cuda-10.2/lib64
sudo chmod a+r /usr/local/cuda-10.2/include/cudnn.h 
sudo chmod a+r /usr/local/cuda-10.2/lib64/libcudnn*

除此之外还需要将其余三个包安装

sudo dpkg -i libcudnn8_8.2.4.15-1+cuda10.2_amd64.deb
sudo dpkg -i libcudnn8-dev_8.2.4.15-1+cuda10.2_amd64.deb
sudo dpkg -i libcudnn8-samples_8.2.4.15-1+cuda10.2_amd64.deb

使用如下命令查看版本，如果有其余多的版本使用sudo apt-get remove libcudnn7、sudo apt-get remove libcudnn7-dev、sudo apt-get remove libcudnn7-samples依次删除

dpkg -l | grep cudnn

ii  libcudnn8                                                    8.2.4.15-1+cuda10.2                             amd64        cuDNN runtime libraries
ii  libcudnn8-dev                                                8.2.4.15-1+cuda10.2                             amd64        cuDNN development libraries and headers
ii  libcudnn8-samples                                            8.2.4.15-1+cuda10.2                             amd64        cuDNN documents and samples

驱动安装

进入官网，选择自己的配置，开始搜索，下载最新版即可


查询是否存在X-server

ps aux | grep X

如果显示如下，说明具有X-server

root     18714  0.0  0.0 415608 25916 tty7     Ssl+ Dec28   0:04 /usr/lib/xorg/Xorg -core :0 -seat seat0 -auth /var/run/lightdm/root/:0 -nolisten tcp vt7 -novtswitch
ubuntu   28400  0.0  0.0  12940  1012 pts/16   S+   03:57   0:00 grep --color=auto X

关闭X的方式有多种，全部输入即可：

sudo service lightdm stop
sudo service gdm stop
sudo service kdm stop

再次查询

ps aux | grep X

出现

ubuntu    5217  0.0  0.0  12940   928 pts/16   S+   04:05   0:00 grep --color=auto X

关闭成功

安装

sudo sh NVIDIA-Linux-x86_64-495.46.run

跟着提示默认点击就行
安装完成后输入nvidia-smi，出现如下画面

Fri Dec 31 04:09:03 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 495.46       Driver Version: 495.46       CUDA Version: 11.5     |
|-------------------------------+----------------------+----------------------+
| 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 ...  Off  | 00000000:02:00.0 Off |                  N/A |
| 19%   30C    P0    58W / 250W |      0MiB / 11178MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  NVIDIA GeForce ...  Off  | 00000000:03:00.0 Off |                  N/A |
| 21%   37C    P0    58W / 250W |      0MiB / 11178MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   2  NVIDIA GeForce ...  Off  | 00000000:82:00.0 Off |                  N/A |
| 17%   32C    P0    57W / 250W |      0MiB / 11178MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   3  NVIDIA GeForce ...  Off  | 00000000:83:00.0 Off |                  N/A |
| 18%   34C    P0    56W / 250W |      0MiB / 11178MiB |      1%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

PyTorch安装

选择适合cuda-10.2的版本

conda create -n pytorch python=3.7
conda activate pytorch
pip install torch torchvision torchaudio

Python 3.7.11 (default, Jul 27 2021, 14:32:16)
[GCC 7.5.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True
>>>

PaddlePaddle安装

**由于之前的环境已经安装nccl，所以我直接支持多卡，更多安装参见paddlepaddle官网

conda create -n paddle python=3.7
conda activate paddle
pip install paddlepaddle-gpu==2.2.1 -i https://mirror.baidu.com/pypi/simple

Python 3.7.11 (default, Jul 27 2021, 14:32:16)
[GCC 7.5.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import paddle
>>> paddle.utils.run_check()
Running verify PaddlePaddle program ...
W1231 04:38:06.920902 20825 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 6.1, Driver API Version: 11.5, Runtime API Version: 10.2
W1231 04:38:06.922672 20825 device_context.cc:465] device: 0, cuDNN Version: 8.3.
PaddlePaddle works well on 1 GPU.
W1231 04:38:08.923141 20825 parallel_executor.cc:617] Cannot enable P2P access from 0 to 2
W1231 04:38:08.923198 20825 parallel_executor.cc:617] Cannot enable P2P access from 0 to 3
W1231 04:38:09.725281 20825 parallel_executor.cc:617] Cannot enable P2P access from 1 to 2
W1231 04:38:09.725327 20825 parallel_executor.cc:617] Cannot enable P2P access from 1 to 3
W1231 04:38:09.725330 20825 parallel_executor.cc:617] Cannot enable P2P access from 2 to 0
W1231 04:38:09.725334 20825 parallel_executor.cc:617] Cannot enable P2P access from 2 to 1
W1231 04:38:10.782135 20825 parallel_executor.cc:617] Cannot enable P2P access from 3 to 0
W1231 04:38:10.782181 20825 parallel_executor.cc:617] Cannot enable P2P access from 3 to 1
W1231 04:38:14.150104 20825 fuse_all_reduce_op_pass.cc:76] Find all_reduce operators: 2. To make the speed faster, some all_reduce ops are fused during training, after fusion, the number of all_reduce ops is 2.
PaddlePaddle works well on 4 GPUs.
PaddlePaddle is installed successfully! Let's start deep learning with PaddlePaddle now.
>>>

TensorRT安装

点击此处去官网下载适配的版本，尽量选择TensorRT 8的版本，使用tar包方式安装，如下所示

按照官方安装例程，根据自己的系统和版本编辑变量，并依次输出如下命令

version="8.2.2.1"
arch="x86_64"
cuda="cuda-10.2"
cudnn="cudnn8.2"
tar xzvf TensorRT-${version}.Linux.${arch}-gnu.${cuda}.${cudnn}.tar.gz

查看解压文件

ls TensorRT-${version}

bin  data  doc  graphsurgeon  include  lib  onnx_graphsurgeon  python  samples  targets  TensorRT-Release-Notes.pdf  uff

拷贝至/usr/local/cuda-10.2/lib64，当然也可以为这个目录新增环境变量export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:

sudo cp -r TensorRT-${version}/lib/* /usr/local/cuda-10.2/lib64

安装TensorRT

cd TensorRT-${version}/python
pip install tensorrt-*-cp37-none-linux_x86_64.whl

安装uff，使用TensorRT和TensorFlow

cd TensorRT-${version}/uff
pip install uff-0.6.9-py2.py3-none-any.whl

使用如下命令验证安装

which convert-to-uff

/home/ubuntu/anaconda3/envs/pytorch/bin/convert-to-uff

安装graphsurgeon

cd TensorRT-${version}/graphsurgeon
pip install graphsurgeon-0.4.5-py2.py3-none-any.whl

安装onnx-graphsurgeon

cd TensorRT-${version}/onnx_graphsurgeon
pip install onnx_graphsurgeon-0.3.12-py2.py3-none-any.whl

验证

cd TensorRT-${version}/samples/sampleMNIST
make
cd ../../bin
./sample_mnist

最终出现如下结果说明安装成功

&&&& RUNNING TensorRT.sample_mnist [TensorRT v8202] # ./sample_mnist
[12/31/2021-07:28:34] [I] Building and running a GPU inference engine for MNIST
[12/31/2021-07:28:36] [I] [TRT] [MemUsageChange] Init CUDA: CPU +160, GPU +0, now: CPU 162, GPU 215 (MiB)
[12/31/2021-07:28:36] [I] [TRT] [MemUsageSnapshot] Begin constructing builder kernel library: CPU 162 MiB, GPU 215 MiB
[12/31/2021-07:28:36] [I] [TRT] [MemUsageSnapshot] End constructing builder kernel library: CPU 183 MiB, GPU 215 MiB
[12/31/2021-07:28:37] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 10.2.3 but loaded cuBLAS/cuBLAS LT 10.2.2
[12/31/2021-07:28:37] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +115, GPU +46, now: CPU 300, GPU 261 (MiB)
[12/31/2021-07:28:37] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +119, GPU +62, now: CPU 419, GPU 323 (MiB)
[12/31/2021-07:28:37] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[12/31/2021-07:28:42] [I] [TRT] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[12/31/2021-07:28:42] [I] [TRT] Detected 1 inputs and 1 output network tensors.
[12/31/2021-07:28:42] [I] [TRT] Total Host Persistent Memory: 5440
[12/31/2021-07:28:42] [I] [TRT] Total Device Persistent Memory: 0
[12/31/2021-07:28:42] [I] [TRT] Total Scratch Memory: 0
[12/31/2021-07:28:42] [I] [TRT] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 1 MiB, GPU 16 MiB
[12/31/2021-07:28:42] [I] [TRT] [BlockAssignment] Algorithm ShiftNTopDown took 0.084057ms to assign 3 blocks to 11 nodes requiring 57857 bytes.
[12/31/2021-07:28:42] [I] [TRT] Total Activation Memory: 57857
[12/31/2021-07:28:42] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 10.2.3 but loaded cuBLAS/cuBLAS LT 10.2.2
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 591, GPU 397 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +10, now: CPU 591, GPU 407 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in building engine: CPU +0, GPU +4, now: CPU 0, GPU 4 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init CUDA: CPU +0, GPU +0, now: CPU 592, GPU 369 (MiB)
[12/31/2021-07:28:42] [I] [TRT] Loaded engine size: 1 MiB
[12/31/2021-07:28:42] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 10.2.3 but loaded cuBLAS/cuBLAS LT 10.2.2
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 592, GPU 379 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 592, GPU 387 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +1, now: CPU 0, GPU 1 (MiB)
[12/31/2021-07:28:42] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 10.2.3 but loaded cuBLAS/cuBLAS LT 10.2.2
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 568, GPU 379 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 568, GPU 387 (MiB)
[12/31/2021-07:28:42] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +0, now: CPU 0, GPU 1 (MiB)
[12/31/2021-07:28:42] [I] Input:
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@.*@@@@@@@@@@
@@@@@@@@@@@@@@@@.=@@@@@@@@@@
@@@@@@@@@@@@+@@@.=@@@@@@@@@@
@@@@@@@@@@@% #@@.=@@@@@@@@@@
@@@@@@@@@@@% #@@.=@@@@@@@@@@
@@@@@@@@@@@+ *@@:-@@@@@@@@@@
@@@@@@@@@@@= *@@= @@@@@@@@@@
@@@@@@@@@@@. #@@= @@@@@@@@@@
@@@@@@@@@@=  =++.-@@@@@@@@@@
@@@@@@@@@@       =@@@@@@@@@@
@@@@@@@@@@  :*## =@@@@@@@@@@
@@@@@@@@@@:*@@@% =@@@@@@@@@@
@@@@@@@@@@@@@@@% =@@@@@@@@@@
@@@@@@@@@@@@@@@# =@@@@@@@@@@
@@@@@@@@@@@@@@@# =@@@@@@@@@@
@@@@@@@@@@@@@@@* *@@@@@@@@@@
@@@@@@@@@@@@@@@= #@@@@@@@@@@
@@@@@@@@@@@@@@@= #@@@@@@@@@@
@@@@@@@@@@@@@@@=.@@@@@@@@@@@
@@@@@@@@@@@@@@@++@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@

[12/31/2021-07:28:42] [I] Output:
0:
1:
2:
3:
4: **********
5:
6:
7:
8:
9:

&&&& PASSED TensorRT.sample_mnist [TensorRT v8202] # ./sample_mnist

备注

注意到：如果使用wget无法链接nvidia下载，修改如下配置

sudo chmod 777 /etc/resolv.conf
vim /etc/resolv.conf

nameserver 8.8.8.8 #google域名服务器
nameserver 8.8.4.4 #google域名服务器

下载完后记得注释掉，以免使用sudo出现sudo: unable to resolve host XXXX