fix inconsistency w/ recent pytorch cuda device logic

Summary: Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/892

Differential Revision: D18109685

Pulled By: jma127

fbshipit-source-id: f96e1080a5577b8ee0748dfdd956bf72bed47474

fairseq/trainer.py

@@ -330,7 +330,7 @@ class Trainer(object):
                     print(msg, file=sys.stderr)
                     if torch.cuda.is_available() and hasattr(torch.cuda, "memory_summary"):
                         for device_idx in range(torch.cuda.device_count()):
-                            print(torch.cuda.memory_summary(device=torch.cuda.device(device_idx)),
+                            print(torch.cuda.memory_summary(device=device_idx),
                                   file=sys.stderr)
                     sys.stderr.flush()