Multi-batch infer results with the same input are different and only the first batch is wrong when running on the DLA

Robotics & Edge Computing Jetson & Embedded Systems Jetson Xavier NX
1

JetPack Version : 4.4

The one-batch model running on the DLA in fp16 always comes out the wrong result, but the multi-batch infer result is different with the same input and only the first batch is wrong.
The model using one or more batches running normally on GPU.

The onnx model is attatched below:
https://drive.google.com/file/d/1jTuSd281-buxbf77sJb5YM3kdz8UQC6z/view?usp=sharing

3

Hi,

Could you also share the inference source code with us for checking?
Thanks.

4

Hi,
The code and model are attached below：
https://drive.google.com/drive/folders/10irumIjOX69sbl0ay2QvgHGJ90CzVZnD?usp=sharing

5

Hi,

Thanks for your sharing.
Will let you know for any progress.

Thanks.

6

OK, waiting for your reply.
Thanks.

7

Hi,

This issue can be reproduced in our environment.
We are passing this to our internal team for further suggestion.

Will share you more information once we got any feedback.

9

OK.
Thanks.

10

Hi,

This issue is fixed in our future release.
Will update here once it releases.

Thanks.

11

OK.
Hope to you release it soon.
Thanks.

13

Hi,

Thanks for your patience.

This issue is fixed in our latest JetPack 4.6 release (TensorRT 8.0).
Please upgrade your environment to avoid this issue.

Thanks.

15

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