OOM when creating TensorRT engine on AGX Xavier using trtexec

Continuing the discussion from Device memory is insufficient to use tactic:

In the thread mentioned above, there is currently an issue on the AGX Xavier devices with their latest TensorRT version. I am running into the same issue. The solution given is simply “wait” for a new version of Jetpack and TensorRT. Could we get more clarity if the issue is? Is the issue for the particular ONNX model, is it trtexec, is it TensorRT? I am asking in case there is another way to create a TensorRT engine on the AGX Xavier such that this isn’t a blocker for development. I could even change the ONNX model if someone could divulge a bit more info as to what is causing this issue. @AastaLLL


Unfortunately, the limitation comes from the TensorRT library.

There is a memory limitation for the Jetson platform due to some internal reason.
It is removed from TensorRT 8.2, which should be included in the next JetPack release.


1 Like