[TorchGeneratorAgent] Batchify default to not sorting

parlai/agents/hred/hred.py

@@ -128,6 +128,10 @@ def batchify(self, obs_batch, sort=True):
 
         Store history vec as context_vec.
         """
+        # NOTE: `sort` is set to True here (Default is False in TorchGeneratorAgent)
+        # TODO: Sorting the batch will result in various local metrics being broadcasted
+        # back to individual examples in the wrong order, such as the lengths of
+        # the context and labels. Aggregate metric reports will still be accurate.
         batch = super().batchify(obs_batch, sort)
         # sum here is list concat, not addition
         context_vec, hist_lens_ = self._pad_tensor(

parlai/agents/seq2seq/seq2seq.py

@@ -209,6 +209,9 @@ def batchify(self, *args, **kwargs):
         Override batchify options for seq2seq.
         """
         kwargs['sort'] = True  # need sorted for pack_padded
+        # TODO: Sorting the batch will result in various local metrics being broadcasted
+        # back to individual examples in the wrong order, such as the lengths of
+        # the context and labels. Aggregate metric reports will still be accurate.
         return super().batchify(*args, **kwargs)
 
     def state_dict(self):

parlai/core/torch_generator_agent.py

@@ -639,7 +639,7 @@ def vectorize(self, *args, **kwargs):
         kwargs['add_end'] = True  # we do want this
         return super().vectorize(*args, **kwargs)
 
-    def batchify(self, obs_batch, sort=True):
+    def batchify(self, obs_batch, sort=False):
         batch = super().batchify(obs_batch, sort=sort)
         if (
             self.beam_block_full_context