search

NVIDIA / TensorRT

NVIDIA® TensorRT™ is an SDK for high-performance deep learning inference on NVIDIA GPUs. This repository contains the open source components of TensorRT.
https://developer.nvidia.com/tensorrt
Apache License 2.0
10.6k stars 2.11k forks source link

CudnnDeconvolution has no valid tactics for this config, skipping. #2280

Closed wywywy01 closed 2 years ago

wywywy01 commented 2 years ago

Description

We get the model onnx file by pytorch-quatization api, and use trtexec to convert int8 engine, met errors as blow, can you give me some help?

Have done, not work: (1) use large workspace

** Error Code 10: Internal Error (Could not find any implementation for node backbone.ida_up.up_1.weight + QuantizeLinear_629_quantize_scale_node + ConvTranspose_631.)

Environment

jpack 4.6.1 as blow

image

Relevant Files

https://drive.google.com/file/d/1gCWO64Kl5eJ70GUfCVpXJtx8hMCeQ5KP/view?usp=sharing

Steps To Reproduce

** trtexec --onnx=demo.onnx --saveEngine=demo.engine --workspace=40240 --int8 --verbose

zerollzeng commented 2 years ago

I can't open the file link, can you upload it to the issue or Google Drive?

zerollzeng commented 2 years ago

if the model is too large then I think prune it and leave only a small netwok contains backbone.ida_up.up_1.weight + QuantizeLinear_629_quantize_scale_node + ConvTranspose_631 will also reproduce your error

wywywy01 commented 2 years ago

I can't open the file link, can you upload it to the issue or Google Drive?

update

zerollzeng commented 2 years ago

I can reproduce this in TRT 8.4 but get passed in TRT 8.5, so it should be fixed in our dev branch, please wait for the next release. @ttyio is this a known issue?

wywywy01 commented 2 years ago

I can reproduce this in TRT 8.4 but get passed in TRT 8.5, so it should be fixed in our dev branch, please wait for the next release. @ttyio is this a known issue?

thanks a lot，I'm curious. Can you explain it further？ Can I make some configuration changes in the old version(8.2) and try？

ttyio commented 2 years ago

We added a new convTranspose backend in 8.5, it means that we will have a boarder coverage for the problem sizes.

@wywywy01 we need upgrade TRT. Otherwise you have to modify your quantization script to remove the Q/DQ before the failed convTranspose. Thanks!

wywywy01 commented 2 years ago

We added a new convTranspose backend in 8.5, it means that we will have a boarder coverage for the problem sizes.

@wywywy01 we need upgrade TRT. Otherwise you have to modify your quantization script to remove the Q/DQ before the failed convTranspose. Thanks!

thanks