TensorRT 8.0: Compiling pytorch 1.8 with CUDA11.3

deployment
1

Hello,

I’m trying to setup TensorRT 8.0, the Jetson TX2 I’m using uses ARM, hence I need to run TensorRT8.0+CUDA11.3. To make sure training and deployment have similar environment, I followed the documentation for TensorRT setup and downloaded CUDA11.3.

Now I’m trying to get pytorch 1.8, given pytorch 1.8 is compiled with CUDA 11.1, I’m trying to compile myself. I got the pytorch 1.8 from: pytorch/releases/tag/v1.8.1 and is trying to follow this guide to compile. However, I can’t get pass the ‘git submodule update --init --recursive’, it says ‘fatal: not a git repository’

Hence:

  1. Should i be following this guide at all? Given im compiling version 1.8
  2. If ‘Yes’ to 1, how do I circumvent the ‘fatal: not a git repository’ error?
  3. if ‘No’ to 1, what should I be doing to compile pytorch 1.8?

Thanks :slight_smile:

2

I can’t put more than 2 links

I got pytorch 1.8 from: Release Small bug fix release · pytorch/pytorch · GitHub

3
  1. Yes, but I would recommend to use the latest release (1.9.0) or the master instead of 1.8.1 (or do you have a specific reason to stick to this older version?)
  2. It seems you are not in the pytorch folder so cd pytorch after executing git clone.