ubuntu18.04 pytorch tensorrt安装

1.tensorrt版本:TensorRT-7.2.1.6，链接https://developer.nvidia.com/nvidia-tensorrt-7x-download

最后把tensorrt文件下的lib和include，cp到/usr/lib和/usr/lib/include下

sudo cp -r ./lib/* /usr/lib

sudo cp -r ./include/* /usr/include

2.cuda版本:cuda10.2,命令行：wget https://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run

如果nvdia驱动不匹配的话，重新装一下驱动：

a.

(1)禁用系统默认显卡驱动：sudo gedit /etc/modprobe.d/blacklist.conf

(2)将下列代码填入文件末尾:

blacklist vga16fb

blacklist nouveau

blacklist rivafb

blacklist rivatv

blacklist nvidiafb

(3)保存，然后更新initramfs:

sudo update-initramfs -u

(4)重启电脑：

sudo reboot

(5)查看是否禁用成功，无输出则禁用成功

lsmod | grep nouveau

b.开始重新安装驱动：

sudo service lightdm stop

sudo apt-get install nvidia-driver-450

sudo service lightdm start

3.cudnn版本:cudnn.8.0.2,链接：https://developer.nvidia.com/rdp/cudnn-archive#a-collapse802-102

4.onnx-tensorrt版本：https://github.com/onnx/onnx-tensorrt.git

mkdir build

cd build

cmake .. -DTENSORRT_ROOT=/mnt/data/hks/tensorrt/TensorRT-7.2.1.6

#其中/mnt/data/hks/tensorrt/TensorRT-7.2.1.6是tensorrrt安装的路径

make -j32

sudo make install

5.还必须安装pycuda，直接pip3 install pucuda会报错，所以，直接到网站：https://pypi.org/project/pycuda/#files，下载源码，然后执行下面的命令：

python3.6 configure.py --cuda-root=/usr/local/cuda
sudo make install

6.安装torch2trt:https://github.com/NVIDIA-AI-IOT/torch2trt