Fix multi-gpu training (fixes #1088)

Summary: Pull Request resolved: https://github.com/pytorch/fairseq/pull/1089

Differential Revision: D17108918

Pulled By: myleott

fbshipit-source-id: 818c77a5bbf3b146028991aca64d79b93f144b28

train.py

@@ -117,10 +117,6 @@ def train(args, trainer, task, epoch_itr):
     valid_subsets = args.valid_subset.split(',')
     max_update = args.max_update or math.inf
     for i, samples in enumerate(progress, start=epoch_itr.iterations_in_epoch):
-        samples = [s for s in samples if len(s) > 0]
-        if len(samples) == 0:
-            continue
-
         log_output = trainer.train_step(samples)
         if log_output is None:
             continue