Clarify data_parallel implementation. (#2488)

- It appears we're unnecessarily replicating the model to the source device, but this is not the case.
- Silence some deprecation warnings that were missed previously.

Author: brendan-ai2

allennlp/training/util.py

@@ -237,7 +237,10 @@ def data_parallel(batch_group: List[TensorDict],
              for batch, device in zip(batch_group, cuda_devices)]
 
     used_device_ids = cuda_devices[:len(moved)]
+    # Counterintuitively, it appears replicate expects the source device id to be the first element
+    # in the device id list. See torch.cuda.comm.broadcast_coalesced, which is called indirectly.
     replicas = replicate(model, used_device_ids)
+
     # We pass all our arguments as kwargs. Create a list of empty tuples of the
     # correct shape to serve as (non-existent) positional arguments.
     inputs = [()] * len(batch_group)

pytest.ini

@@ -28,5 +28,8 @@ filterwarnings =
     ignore:@asynchronous is deprecated, use coroutines instead::tornado\.web
     #
     ignore:encoding is deprecated, Use raw=False instead.:PendingDeprecationWarning:msgpack_numpy
+    # For `spacy==2.0.11`
+    ignore:Direct calling implementation's unpack.*:PendingDeprecationWarning:msgpack_numpy
+    ignore:The binary mode of fromstring is deprecated.*:DeprecationWarning:msgpack_numpy
     # 4. ignore these `ExperimentalFeatureWarning` for now, but record them once
     once:This particular transformer implementation is a provisional feature.*::allennlp\.modules\.seq2seq_encoders\.bidirectional_language_model_transformer