Output reproducibility issue while converting Mvitv2 torch model to ONNX format

We tried to convert Mvitv2 torch model to onnx using torch.onnx.export.Even thought the input is same , we are getting 2 different outputs for torch and onnx models (which are way different). However on the dummy input we used while conversion its producing same output for torch and onnx . How can convert our Mvitv2 torch model to onnx such that the same results are reproduced ?