Differences of CPU and GPU in exporting onnx

[Macsim2]

Hi, I'm execute onnx model usefully thanks to export_onnx_cpu or gpu python file.
But Sometime I face this situation that model output differently between CPU onnx model and GPU onnx model with big differences. I mean that exported CPU onnx model is right, but GPU onnx model could output wrong result always although two models have same options(.pt file, train.yaml, cmvn file, chunk size, left chunks .. etc)

So, I'm curious that has anyone else seen this problem and What is the cause of the problem
If you guys know anything about this situation, I would appreciate it, if you could let me know