Int8 calculation problem

[CoinCheung]

Hi,

I posted the error message in the repo of Tensorrt and they refered to this repo, so I open an issue here. The problem is that when I quantize the model in pytorch with modelopt and export it to onnx, the tensorrt will fail to compile the onnx file into tensorrt engine..

Here is the link with an example code piece which shows how to reproduce the error message:

NVIDIA/TensorRT#4095 (comment)

Please help me to make it work.

[cjluo-omniml]

Have you tried the onnx PTQ workflow which exports to the onnx first then do the quantization? See https://github.com/NVIDIA/TensorRT-Model-Optimizer/tree/main/onnx_ptq