Sample instead of greedy decoding by default in generate

src/transformers/modeling_utils.py

@@ -586,7 +586,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         self,
         input_ids=None,
         max_length=None,
-        do_sample=None,
+        do_sample=True,
         num_beams=None,
         temperature=None,
         top_k=None,
@@ -617,7 +617,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                 The max length of the sequence to be generated.  Between 1 and infinity. Default to 20.
 
             do_sample: (`optional`) bool
-                If set to `False` greedy decoding is used. Otherwise sampling is used. Default to greedy sampling.
+                If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `True`.
 
             num_beams: (`optional`) int
                 Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.