Missing tensorrt binaries in nvidia-tensorrt package from jetpack r35.4 repository

We face two issues with the new release of jetpack 5.1.2:

  • The previous l4t-base container 35.3.1 is replaced by the new release 35.4.1. Both containers have an identical checksum and use the same r35.4 repository by default

  • The structure of nvidia-tensorrt package from the new jetpack r35.4 is changed and the new package only depends on tensorrt-libs. When installing the new package, tensorrt is not actually installed and binaries (tensorrt and trtexec) are missing. But the nvidia-tensorrt-dev package is fine as it installs the tensorrt package

nvidia-tensorrt r35.3 dependencies:

libnvinfer-bin (= 8.5.2-1+cuda11.4), libnvinfer8 (= 8.5.2-1+cuda11.4), libnvparsers8 (= 8.5.2-1+cuda11.4), libnvinfer-plugin8 (= 8.5.2-1+cuda11.4), libnvonnxparsers8 (= 8.5.2-1+cuda11.4)

nvidia-tensorrt r35.4 dependencies:

tensorrt-libs (= 8.5.2.2-1+cuda11.4)

tensorrt-libs:
libnvinfer8 (= 8.5.2-1+cuda11.4), libnvinfer-plugin8 (= 8.5.2-1+cuda11.4), libnvparsers8 (= 8.5.2-1+cuda11.4), libnvonnxparsers8 (= 8.5.2-1+cuda11.4)

Could you please verify the issues?

1 Like

Hi,

Yes, the packaging way does change.
When you install tensorrt, it should also install the libnvinfer-bin and libnvinfer-samples.

Thanks.

Thank you for the response. Could you also clarify the first issue regarding the l4t-base container tag 35.3.1, is this change on purpose?

As the container size is important for us, the installation of tensorrt package is not feasible. Based on the jetpack documentation the nvidia-tensorrt is supposed to be a minimal and runtime only installation of tensorrt, then what would be the alternative in the new repository r35.4?

Hi,

We need to check with our internal team to get more details.
Will let you know the following.

Thanks.

Hello @AastaLLL


As you can see, the l4t-base:35.4.1 and l4t-base:35.3.1 have the very same Digest Hashsum:
sha256:7664d384942fbd27c9ea450c62525606848522540084febd110617d91cacd6f8

Our CI pipeline uses l4t-base:35.3.1 and our most recent release was broken because the apt sources in that image are incorrect. This has been and is still currently delaying customer projects for us.

We have manually fixed the issue in our builds by changing the apt sources - but we need to be able to rely on the NGC images to be correct going into the future. Can we get verification of this issue and also feedback from Nvidia on how this can be avoided in the future. Thanks.

Hi, both

Sorry about this issue.

It seems that the r35.3.1 is mis-updated by the r35.4.1 image.
We are working on this internally. Will get back to you soon.

Thanks.

2 Likes

Hello @AastaLLL,
it has been a week. Is there any update on the matter?

As we have mentioned earlier, we rely on the integrity of the NGC for our CI/CD pipelines. Can you elaborate on the status of the resolution and indicate a resolution date?

Moreover, is this the right place to address this issue or should we open a ticket somewhere else?

Thanks in advance.

Hi,

We have filed an internal bug to our internal team.
Will check the latest status and get back to you.

Thanks.

Hi,

Issue should be fixed.
Please help to confirm.

Thanks.

Thank you for the effort. The issue with the l4t-base:35.3.1 image is resolved.

For anyone who is encountering the missing tensorrt binaries after installing nvidia-tensorrt from r35.4 repository, as the structure of this meta package has been changed, in addition to nvidia-tensorrt, you also need to install libnvinfer-bin to get the minimal installation of tensorrt like prior to jetpack 5.1.2.

Hi,

Thanks for the confirmation.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.