Pipeline: use tokenizer pad token at generation time if the model pad token is unset. (#29614)

src/transformers/pipelines/automatic_speech_recognition.py

@@ -311,14 +311,14 @@ class AutomaticSpeechRecognitionPipeline(ChunkPipeline):
 
         forward_params = defaultdict(dict)
         if max_new_tokens is not None:
-            forward_params["generate_kwargs"]["max_new_tokens"] = max_new_tokens
+            forward_params["max_new_tokens"] = max_new_tokens
         if generate_kwargs is not None:
             if max_new_tokens is not None and "max_new_tokens" in generate_kwargs:
                 raise ValueError(
                     "`max_new_tokens` is defined both as an argument and inside `generate_kwargs` argument, please use"
                     " only 1 version"
                 )
-            forward_params["generate_kwargs"].update(generate_kwargs)
+            forward_params.update(generate_kwargs)
 
         postprocess_params = {}
         if decoder_kwargs is not None:
@@ -456,10 +456,7 @@ class AutomaticSpeechRecognitionPipeline(ChunkPipeline):
                 processed["stride"] = stride
             yield {"is_last": True, **processed, **extra}
 
-    def _forward(self, model_inputs, return_timestamps=False, generate_kwargs=None):
-        if generate_kwargs is None:
-            generate_kwargs = {}
-
+    def _forward(self, model_inputs, return_timestamps=False, **generate_kwargs):
         attention_mask = model_inputs.pop("attention_mask", None)
         stride = model_inputs.pop("stride", None)
         is_last = model_inputs.pop("is_last")

src/transformers/pipelines/base.py

@@ -885,6 +885,16 @@ class Pipeline(_ScikitCompat):
         self._num_workers = kwargs.pop("num_workers", None)
         self._preprocess_params, self._forward_params, self._postprocess_params = self._sanitize_parameters(**kwargs)
 
+        # Pipelines calling `generate`: if the tokenizer has a pad token but the model doesn't, set it in the
+        # forward params so that `generate` is aware of the pad token.
+        if (
+            self.tokenizer is not None
+            and self.model.can_generate()
+            and self.tokenizer.pad_token_id is not None
+            and self.model.generation_config.pad_token_id is None
+        ):
+            self._forward_params["pad_token_id"] = self.tokenizer.pad_token_id
+
         if self.image_processor is None and self.feature_extractor is not None:
             if isinstance(self.feature_extractor, BaseImageProcessor):
                 # Backward compatible change, if users called

src/transformers/pipelines/conversational.py

@@ -196,9 +196,7 @@ class Conversation:
     build_pipeline_init_args(has_tokenizer=True),
     r"""
         min_length_for_response (`int`, *optional*, defaults to 32):
-            The minimum length (in number of tokens) for a response.
-        minimum_tokens (`int`, *optional*, defaults to 10):
-            The minimum length of tokens to leave for a response.""",
+            The minimum length (in number of tokens) for a response.""",
 )
 class ConversationalPipeline(Pipeline):
     """
@@ -241,17 +239,13 @@ class ConversationalPipeline(Pipeline):
         if self.tokenizer.pad_token_id is None:
             self.tokenizer.pad_token = self.tokenizer.eos_token
 
-    def _sanitize_parameters(
-        self, min_length_for_response=None, minimum_tokens=None, clean_up_tokenization_spaces=None, **generate_kwargs
-    ):
+    def _sanitize_parameters(self, min_length_for_response=None, clean_up_tokenization_spaces=None, **generate_kwargs):
         preprocess_params = {}
         forward_params = {}
         postprocess_params = {}
 
         if min_length_for_response is not None:
             preprocess_params["min_length_for_response"] = min_length_for_response
-        if minimum_tokens is not None:
-            forward_params["minimum_tokens"] = minimum_tokens
 
         if "max_length" in generate_kwargs:
             forward_params["max_length"] = generate_kwargs["max_length"]
@@ -304,7 +298,7 @@ class ConversationalPipeline(Pipeline):
             input_ids = tf.constant([input_ids])
         return {"input_ids": input_ids, "conversation": conversation}
 
-    def _forward(self, model_inputs, minimum_tokens=10, **generate_kwargs):
+    def _forward(self, model_inputs, **generate_kwargs):
         n = model_inputs["input_ids"].shape[1]
         conversation = model_inputs.pop("conversation")
         if "max_length" not in generate_kwargs and "max_new_tokens" not in generate_kwargs:

src/transformers/pipelines/document_question_answering.py

@@ -419,14 +419,14 @@ class DocumentQuestionAnsweringPipeline(ChunkPipeline):
                     "is_last": span_idx == num_spans - 1,
                 }
 
-    def _forward(self, model_inputs):
+    def _forward(self, model_inputs, **generate_kwargs):
         p_mask = model_inputs.pop("p_mask", None)
         word_ids = model_inputs.pop("word_ids", None)
         words = model_inputs.pop("words", None)
         is_last = model_inputs.pop("is_last", False)
 
         if self.model_type == ModelType.VisionEncoderDecoder:
-            model_outputs = self.model.generate(**model_inputs)
+            model_outputs = self.model.generate(**model_inputs, **generate_kwargs)
         else:
             model_outputs = self.model(**model_inputs)
 

src/transformers/pipelines/image_to_text.py

@@ -74,7 +74,7 @@ class ImageToTextPipeline(Pipeline):
         )
 
     def _sanitize_parameters(self, max_new_tokens=None, generate_kwargs=None, prompt=None, timeout=None):
-        forward_kwargs = {}
+        forward_params = {}
         preprocess_params = {}
 
         if prompt is not None:
@@ -82,18 +82,17 @@ class ImageToTextPipeline(Pipeline):
         if timeout is not None:
             preprocess_params["timeout"] = timeout
 
-        if generate_kwargs is not None:
-            forward_kwargs["generate_kwargs"] = generate_kwargs
         if max_new_tokens is not None:
-            if "generate_kwargs" not in forward_kwargs:
-                forward_kwargs["generate_kwargs"] = {}
-            if "max_new_tokens" in forward_kwargs["generate_kwargs"]:
+            forward_params["max_new_tokens"] = max_new_tokens
+        if generate_kwargs is not None:
+            if max_new_tokens is not None and "max_new_tokens" in generate_kwargs:
                 raise ValueError(
-                    "'max_new_tokens' is defined twice, once in 'generate_kwargs' and once as a direct parameter,"
-                    " please use only one"
+                    "`max_new_tokens` is defined both as an argument and inside `generate_kwargs` argument, please use"
+                    " only 1 version"
                 )
-            forward_kwargs["generate_kwargs"]["max_new_tokens"] = max_new_tokens
-        return preprocess_params, forward_kwargs, {}
+            forward_params.update(generate_kwargs)
+
+        return preprocess_params, forward_params, {}
 
     def __call__(self, images: Union[str, List[str], "Image.Image", List["Image.Image"]], **kwargs):
         """
@@ -164,7 +163,7 @@ class ImageToTextPipeline(Pipeline):
 
         return model_inputs
 
-    def _forward(self, model_inputs, generate_kwargs=None):
+    def _forward(self, model_inputs, **generate_kwargs):
         # Git model sets `model_inputs["input_ids"] = None` in `preprocess` (when `prompt=None`). In batch model, the
         # pipeline will group them into a list of `None`, which fail `_forward`. Avoid this by checking it first.
         if (
@@ -174,8 +173,6 @@ class ImageToTextPipeline(Pipeline):
         ):
             model_inputs["input_ids"] = None
 
-        if generate_kwargs is None:
-            generate_kwargs = {}
         # FIXME: We need to pop here due to a difference in how `generation.py` and `generation.tf_utils.py`
         #  parse inputs. In the Tensorflow version, `generate` raises an error if we don't use `input_ids` whereas
         #  the PyTorch version matches it with `self.model.main_input_name` or `self.model.encoder.main_input_name`

src/transformers/pipelines/table_question_answering.py

@@ -376,7 +376,7 @@ class TableQuestionAnsweringPipeline(Pipeline):
         inputs["table"] = table
         return inputs
 
-    def _forward(self, model_inputs, sequential=False):
+    def _forward(self, model_inputs, sequential=False, **generate_kwargs):
         table = model_inputs.pop("table")
 
         if self.type == "tapas":
@@ -385,7 +385,7 @@ class TableQuestionAnsweringPipeline(Pipeline):
             else:
                 outputs = self.batch_inference(**model_inputs)
         else:
-            outputs = self.model.generate(**model_inputs)
+            outputs = self.model.generate(**model_inputs, **generate_kwargs)
         model_outputs = {"model_inputs": model_inputs, "table": table, "outputs": outputs}
         return model_outputs