TensorRT 8.5.2 PASS but TensorRT 10.4.0 FAILED

[yijun02]

I am trying to convert an open-clip model to TensorRT:

I converted the model and found that there was a problem with the accuracy so I used the following command:

polygraphy run model.onnx --trt --onnxrt > comparison_results.txt

the command's result:
comparison_results_r35.txt
comparison_results_r36.txt

docker image:
dustynv/l4t-pytorch:2.2-r35.4.1
dustynv/l4t-pytorch:r36.4.0

code and onnx:
https://drive.google.com/file/d/1YXDKsnEYa_v6Un44lG12QVTVw-dzNr5E/view?usp=sharing

[lix19937]

Try to use trt10.6 or trt10.7.

[yijun02]

I tried using trt10.7 but had the same problem.

I used the following command:
polygraphy run model.onnx --trt --onnxrt > comparison_results.txt

the command's result:
comparison_results.txt

docker image:
nvcr.io/nvidia/pytorch:24.12-py3

Is it the same problem as #4333 ?

[yijun02]

I'm using TensorRT 10.8 in the nvcr.io/nvidia/pytorch:25.01-py3 container works as expected. But how do I use TensorRT 10.8 on Jetson without using the PyTorch container?

[kevinch-nv]

Due to some issues a Jetson TensorRT 10.8 is not available on Jetson. Jetson packages will be available again in the 10.9 release. Given that the original accuracy issue has been resolved in 10.8, closing this issue.