How to install tensorrt for pytorch in conda environment

[govindamagrawal]

Hi,
Can someone please guide how to install tensorrt properly in conda environment. I have install tensorrt though downloading .whl file from nvidia repo and it got installed by running this command in the conda activated env:

pip install /home/govindam/TensorRT-7.0.0.11/python/tensorrt-7.0.0.11-cp37-none-linux_x86_64.whl

But it also requires to install other dependencies like python3-libnvinfer-dev etc which I guess can be installed only via sudo, but will not work in the conda environment. I am getting error like this when trying to import torch2trt:

>>> from torch2trt import torch2trt
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/govindam/anaconda3/envs/tf_c9/lib/python3.7/site-packages/torch2trt-0.1.0-py3.7-linux-x86_64.egg/torch2trt/__init__.py", line 1, in <module>
    from .torch2trt import *
  File "/home/govindam/anaconda3/envs/tf_c9/lib/python3.7/site-packages/torch2trt-0.1.0-py3.7-linux-x86_64.egg/torch2trt/torch2trt.py", line 2, in <module>
    import tensorrt as trt
  File "/home/govindam/anaconda3/envs/tf_c9/lib/python3.7/site-packages/tensorrt/__init__.py", line 1, in <module>
    from .tensorrt import *
ImportError: libnvrtc.so.10.0: cannot open shared object file: No such file or directory

Can someone please help to install the package ? Thanks in advance.

[ravijo]

Please follow the instructions given on the official website . Assuming you have installed Cuda and cuDNN properly, you can install TensorRT using pip using the following commands:

pip install nvidia-pyindex
pip install nvidia-tensorrt

Please make sure to have updated versions of pip and setuptools by executing pip3 install --upgrade setuptools pip. After installing TensorRT, you can verify the installation as shown below:

(my_env) ravi@home:~$ ipython
Python 3.7.10 (default, Feb 26 2021, 18:47:35) 
Type 'copyright', 'credits' or 'license' for more information
IPython 7.21.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import tensorrt

In [2]: tensorrt.__version__
Out[2]: '7.2.3.4'

In [3]: assert tensorrt.Builder(tensorrt.Logger())

In [4]: tensorrt.__file__
Out[4]: '/home/ravi/anaconda/envs/my_env/lib/python3.7/site-packages/tensorrt/__init__.py'

In [5]: import torch

In [6]: torch.__version__
Out[6]: '1.7.0'

In [7]: torch.version.cuda
Out[7]: '11.0'

In [8]: ls /home/ravi/anaconda/envs/my_env/lib/python3.7/site-packages/tensorrt/
__init__.py  libmyelin.so.1  libnvinfer_plugin.so.7  libnvinfer.so.7  libnvonnxparser.so.7  libnvparsers.so.7  __pycache__/  tensorrt.so*

You may see the following installation locations too:

• tensorrt.so is installed at /home/ravi/anaconda/envs/my_env/lib/python3.7/site-packages/tensorrt/
• libnvrtc.so is installed at /home/ravi/anaconda/envs/my_env/lib/libnvrtc.so

[MalusJ]

ImportError: libnvrtc.so.10.0: cannot open shared object file: No such file or directory

I think it's because your cuda version doesnt match what TensorRT needed. Try update your cuda to 10.0 and run it.

[govindamagrawal]

Thanks for the reply,

[gioipv]

I installed tensorrt with tar file in conda environment.
in the steps to install tensorrt with tar file, using pip install instead of sudo pip install. Because if u use sudo, the tensorrt use python system instead of python in conda. and u have to update python path to use tensorrt, but it is not the python version in your env.