TensorRT QuantDequantLinearHelper Error: Scale Coefficients Must All Be Positive

[jhcho90]

Thank you for your help.

I used the onnx_ptq example to quantize my own model.
I applied it and successfully obtained quantized results.

However, when I tried to evaluate latency on Orin-X, it failed with the following error:

[E] [TRT] ModelImporter.cpp:776: ERROR: builtin_op_importers.cpp:1197 In function QuantDequantLinearHelper:
[6] Assertion failed: scaleAllPositive && "Scale coefficients must all be positive"

My TensorRT version is 8.6.1, and Orin-X uses the same version.
The opset of ONNX is 21.

Thank you.

[i-riyad]

Please share the input ONNX model and reproducible code/commands. Also can you try TensorRT 10.13.x.