Are some layers implemented in DFINE and DEIM which aren't supported?

[horngjason]

Hi all,

While I was playing with DFINE and DEIM models, I couldn't convert the model from .pth -> ONNX -> TRT INT8. Has anyone met the same situation?

It's successful to convert the model to FP16 or FP8. Yet, once I set INT8 in the trtexec, what I got in the end would be a "FAILED TensorRT.....".

Failed_profile_fp16int8.txt

Possible solutions I tried are:

• Use simplified ONNX model
• Set ONNX model in opset17 instead opset16
• Set allocationStrategy=runtime in trtexec
  Neither of them solved my case.

My environment:

• TensorRT: 10.5, 10.6.0.26 and 10.8.0.43
• CUDA: 11.8
• DFINE model type: dfine_hgnetv2_n_coco.yml

Many thanks!

[kevinch-nv]

Are you able to share your ONNX model? What GPU are you running on?

[horngjason]

Hi @kevinch-nv, thanks for the reply.

I'm testing on a RTX4060 Laptop GPU (on ASUS TUF GAMING A14).

Here are the official model that I used their official converter without simplifying: https://drive.google.com/drive/folders/1vMn1aqMOnS2w_W7FJKNaVbt72TukxfDR?usp=drive_link. They are the smallest two versions of DFINE and are in OPSET17.

I've tried to add arguments such as
● trtexec --onnx=./dfine/opset17/dfine_s_coco.onnx --saveEngine=./dfine/opset17/dfine_s_coco.engine --[Switch between fp16 or int8] --stronglyType --verbose
● trtexec --onnx=./dfine/opset17/dfine_s_coco.onnx --saveEngine=./dfine/opset17/dfine_s_coco_int8.trt --[Switch between fp16 or int8] --precisionConstraints=prefer --layerPrecisions=*:fp16 --verbose

[lix19937]

Maybe a bug, would you like to try the latest version to check if the error still occurs?