Incorrect output when using FP16 input after TensorRT conversion

[zgspose]

I’m running into a strange issue.
Before converting my model to TensorRT, using FP16 for both the input and the model works fine, and the output is correct.
However, after converting the model to TensorRT, FP16 inputs produce incorrect results, while FP32 inputs work as expected.

This behavior is unexpected — is there a known reason why FP16 inputs would cause wrong outputs after TensorRT conversion?

[zavinator]

I haven’t encountered this issue myself. In my workflow, I first convert the model to ONNX and then use NVIDIA’s trtexec.exe to generate the TensorRT engine.