Fixing empty prompts for text-generation when BOS exists.

src/transformers/pipelines/text_generation.py

@@ -158,6 +158,9 @@ def preprocess(self, prompt_text, prefix=""):
 
     def _forward(self, model_inputs, **generate_kwargs):
         input_ids = model_inputs["input_ids"]
+        # Allow empty prompts
+        if input_ids.shape[1] == 0:
+            input_ids = None
         prompt_text = model_inputs.pop("prompt_text")
         generated_sequence = self.model.generate(input_ids=input_ids, **generate_kwargs)  # BS x SL
         return {"generated_sequence": generated_sequence, "input_ids": input_ids, "prompt_text": prompt_text}

tests/test_pipelines_text_generation.py

@@ -106,3 +106,14 @@ def run_pipeline_test(self, model, tokenizer, feature_extractor):
         outputs = text_generator("This is a test", return_full_text=True)
         self.assertEqual(outputs, [{"generated_text": ANY(str)}])
         self.assertTrue(outputs[0]["generated_text"].startswith("This is a test"))
+
+        # Empty prompt is slighly special
+        # it requires BOS token to exist.
+        # Special case for Pegasus which will always append EOS so will
+        # work even without BOS.
+        if text_generator.tokenizer.bos_token_id is not None or "Pegasus" in tokenizer.__class__.__name__:
+            outputs = text_generator("")
+            self.assertEqual(outputs, [{"generated_text": ANY(str)}])
+        else:
+            with self.assertRaises((ValueError, AssertionError)):
+                outputs = text_generator("")