I see three failing tests because of two problems:

1. test_dpr.py calls initialize_device_settings(use_cuda=False)
   therefore, torch.distributed.init_process_group(backend="nccl") is not executed and thus
   rank = torch.distributed.get_rank() results in a RuntimeError

I suggest the following change:

• except (AssertionError, RuntimeError):

2. The probability value of a prediction changed slightly in one of the tests:
   

   FARM/test/test_ner_amp.py

   Line 97 in 816b4e3

   assert result[0]["predictions"][0][0]["probability"] > 0.124

   
   E assert 0.123253375 > 0.124

I suggest the following change:

• assert np.isclose(result[0]["predictions"][0][0]["probability"], 0.124, rtol=0.05)

Hooray! Thanks for this, no longer needing to keep an old version of torch around is really convenient :)