search

onnx / onnx-tensorrt

ONNX-TensorRT: TensorRT backend for ONNX
Apache License 2.0
2.95k stars 545 forks source link

Create temp weights for constants #973

Closed olesalscheider closed 5 months ago

olesalscheider commented 5 months ago

Without this, TRT complains that multiple weights with the same value but different counts are added to the graph (using the dynamo based ONNX export of pytorch 2):

[TRT] [E] [network.cpp::setWeightsName::3797] Error Code 1: Internal Error (Error: Weights of same values but of different counts are used in the network!)

It seems that some of the weights go out of scope and their addresses get reused.

I fixed this by copying the value to tempWeights which will not go out of scope, but I do not know if this is the proper fix.

olesalscheider commented 5 months ago

This PR is not relevant anymore, I documented the underlying issue here: https://github.com/onnx/onnx-tensorrt/issues/974