Missing libnvdla_compiler.so in JetPack 7.2 on Jetson Orin NX

Robotics & Edge Computing Jetson Systems Jetson Orin NX
1

Hello,

I have recently set up a Jetson Orin NX using the newly released JetPack 7.2. However, when I attempted to use trtexec (or run containerized applications), I encountered errors/warnings regarding a missing libnvdla_compiler.so file. Here is the log from nvidia-ctk

Jun 03 06:22:06 orin-nx70 nvidia-ctk[7207]: time="2026-06-03T06:22:06Z" level=warning msg="Could not locate /usr/lib/aarch64-linux-gnu/nvidia/libnvdla_compiler.so: /usr/lib/aarch64-linux-gnu/nvidia/libnvdla_compiler.so: not found"

I checked the host directory, and the compiler library is indeed missing, while other DLA-related libraries exist:

$ ls /usr/lib/aarch64-linux-gnu/nvidia/ -la | grep dla
-rw-r–r-- 1 root root 289512 Jun 1 21:16 libnvcudla.so
-rw-r–r-- 1 root root 7347616 Jun 1 21:16 libnvdla_runtime.so
-rw-r–r-- 1 root root 70000 Jun 1 21:16 libnvmedia_dla.so

I also ran apt-file search to see if there is any package providing this file in the JetPack 7.2 repository, but it returned nothing.

On my other/older Jetson environment, this file used to be provided by the nvidia-l4t-core (or similar nvidia-l4t-dla-compiler) package.
It seems that libnvdla_compiler.so is missing from the apt repository or the BSP package for JetPack 7.2.
Could you please check if this library was omitted from the JetPack 7.2 release or if there is an alternative package I should install?

Thank you.

4

Hi,

Do you need to deploy the model on DLA?
When testing trtexec locally, it can run deploy the model on GPU correctly:

$ ./trtexec --onnx=/home/nvidia/ResNet50.onnx 
&&&& RUNNING TensorRT.trtexec [TensorRT v101602] [b10] # ./trtexec --onnx=/home/nvidia/ResNet50.onnx
...
&&&& PASSED TensorRT.trtexec [TensorRT v101602] [b10] # ./trtexec --onnx=/home/nvidia/ResNet50.onnx

Although we do see some error when switching to --useDLACore=0 --allowGPUFallback which is under inverstigation.

Thanks.

6

Thank you for the update and for confirming that it runs successfully on the GPU.

Yes, we definitely need to deploy this model on the DLA.
We have already optimized this specific model for the DLA in a previous JetPack version. Furthermore, our entire system architecture and performance targets are designed under the strict assumption that this workload runs on the DLA to offload the GPU.

Thank you very much for investigating the error with --useDLACore=0. We highly appreciate your support and look forward to a resolution on this issue.

Best regards,

7

Also seeing this issue on an Orin AGX

8

Hi, all

DLA will be supported in future TensorRT releases.
Thanks and sorry for the inconvenience.