failure to campare trt and onnxrt by int8 when running polygraphy on GPU 3090

[J-xinyu]

Description

1. I've been testing on PyTorch and noticed an AP discrepancy (0.5011xx down to 0.4941xx) between the pth model and the engine, which was generated using trtexec ptq.onnx ptq.engine --int8 --fp16 . I'm hoping to use polygraphy to debug and find out which layer is causing the engine's accuracy to dip.
2. But I see the --trt-outputs mark all isn't working with my QDQ int8 model.
3. So I use polygraphy run ptq.onnx --trt --int8 --onnxrt --fail-fast --atol 1e-2 --rtol 1e-3 to find layers with misaligned accuracy step by step.But I get

[E] 10: [optimizer.cpp::computeCosts::3728] Error Code 10: Internal Error (Could not find any implementation for node img_view_transformer.depth_net.depth_conv.3.aspp3.atrous_conv.weight + /depth_net/depth_conv/depth_conv.3/aspp3/atrous_conv/_weight_quantizer/QuantizeLinear + /depth_net/depth_conv/depth_conv.3/aspp3/atrous_conv/Conv.)

4. when i try polygraphy run ptq.onnx --trt --fp16 --int8

[E] 10: [optimizer.cpp::computeCosts::3728] Error Code 10: Internal Error (Could not find any implementation for node img_view_transformer.depth_net.depth_conv.3.aspp3.atrous_conv.weight + /depth_net/depth_conv/depth_conv.3/aspp3/atrous_conv/_weight_quantizer/QuantizeLinear + /depth_net/depth_conv/depth_conv.3/aspp3/atrous_conv/Conv.)

Environment

TensorRT Version: 8.5.3.1

NVIDIA GPU: 3090

NVIDIA Driver Version: 515.48.07

CUDA Version: 11.6

CUDNN Version: 8.9.6.50

Operating System: Ubuntu18.04

Python Version: 3.8.19

ONNX Version: 1.13.0

PyTorch Version: 1.13.0

polygraphy Version: 0.49.9

Relevant Files

ptq.onnx link: 百度网盘

Steps To Reproduce

Commands or scripts:
polygraphy run ptq.onnx --trt --int8 --onnxrt --fail-fast --atol 1e-2 --rtol 1e-3
polygraphy run ptq.onnx --trt --fp16 --int8

Have you tried the latest release?: No

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt): Yes,its work

[lix19937]

0.5011xx down to 0.4941xx is normal.