TRTorch on Jetson Xavier AGX

Hi Guys,

I’m trying to install the TRTorch on my brand new Xavier AGX, flashed with JetPack 4.4.

The problem is that TRTorch supports pytorch 1.5.1 and we don’t have the pytorch 1.5.1 available for Jetson. I checked the l4t-pytorch docker containers and we just have 1.5.0 and 1.6.0. Both fail, as cuda version is different (1.5.0) and header files changed (for 1.6.0).

Do you know how to install pytorch 1.5.1 in xavier agx ?
Is some change in progress to make it work on the 1.6.0?

Best regards

Sidnei

Hi,

Based on the page, it looks like the TRTorch v0.0.3 release is compatible on JetPack4.4.

Do you mean any error when using it?
Thanks

Hi @AastaLLL,

The TRTorch works on pytorch 1.5.1 only. And I can’t find this version available for Jetson. The docker containers are only 1.5.0 and 1.6.0. Could you assist me how to get the pytorch 1.5.1 working on jetson JetPack 4.4?

Answering your question: the error when building TRTorch is that it can’t find the quantitizer.h header file, which was removed from pytorch 1.5.1 to 1.6.0 version.

Cheers.

Thanks for the clarification.
We are checking with our internal team. Will update more information with you later.

Thanks.

1 Like

Hi,

Please give pyTorch 1.5.0 a try.
Our internal team confirm that pyTorch 1.5.0 should be compatible with TRTorch v0.0.3 as well.

Do you meet any error when using it?
If yes, would you mind to share the log with us?

Thanks.

Thank you @AastaLLL,

I’m using the latest version of JetPack 4.4, with cuda and cudnn installed via the SDK Manager.

However when I try to run:

docker pull nvcr.io/nvidia/l4t-pytorch:r32.4.2-pth1.5-py3

docker run --rm -it --runtime nvidia nvcr.io/nvidia/l4t-pytorch:r32.4.2-pth1.5-py3

python3

import torch

Segmentation fault (core dumped)

I guess I should downgrade my jetpack to run this. Does that make any sense?

Regards
Sidnei

Hi,

JetPack4.4 product release is using r32.4.3 OS image.
So the r32.4.2-pth1.5-py3 cannot work on it.
Please give JetPack 4.4 DP a try, which uses r32.4.2.

The reason for this is that our TRTorch doesn’t support 1.6.0 yet.
We will pass this issue to our internal team.

Thanks.