Clarification about dynamic layers support on DLA core (Jetson AGX Orin 64 GB)

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Orin
, ,
1

Hi,
I’m using a Jetson AGX Orin 64 GB with JetPack 6.0 to generate a TensorRT engine to target a DLA core.

My model contains dynamic outputs which is decided at runtime (for example : It depends on number of person in the image)

I am able to execute the quantized (INT8) model on GPU, while targeting to DLA ,getting below error

[valueCloner.cpp::replace::31] Error Code 1: Internal Error (Tensor is not known to this ValueCloner)

Is there a possible way to fix this issue, or has it been addressed and updated in the latest Jetson package or TensorRT?

image
image1309×398 69.9 KB

3

Hi,

Have you tried if the model can work with the GPU backend?
If GPU can work fine, could you share the ONNX file with us for checking?

Thanks.

4

Hello @AastaLLL,

Yes, the model is running successfully on the hardware using the GPU backend.

Since it is proprietary model, we do not have the authority to share ; however, I can provide the necessary information regarding output logs or some other details.

Let me know if you need any further information required

5

Hi,

This is a known issue, please check the below topic for more info:

However, due to the limited resources, this issue hasn’t been fixed yet.
We have asked our internal team to check it further and will update more with you later.

Thanks.

7

Hi ,

I have uploaded a dummy ONNX model to replicate the error on your side:
onnx_model.zip (36.0 MB)

We are encountering this error during both FP16 and INT8 conversion while targeting DLA.
Please check and provide a solution.

8

Hi,

Thanks for sharing the model.

We tested it with TensorRT 10.7 and it can be inferred on DLA without error.
Please find the below comment to upgrade:

Thanks.

9

Hello @AastaLLL

Thank you for the solution.

We’d like to confirm one thing from your side:

you mentioned you were able to inference the our model on the DLA, but could you please clarify whether it was only on the DLA or in (DLA+GPU) hybrid mode?.