Result is inconsistent with GPU

[henrywoo]

In the test_consistency.py file, if we change the device from cpu to gpu(torch.device("cuda:0")), the jit_probs and py_probs have non-trivial difference.

You can see jit_probs is:

[[0.983 0.01359 0.00349]]

but py_probs is:

[[0.984 0.01278 0.003307]]

Did I do something wrong or Is this a bug or something?