feat[vLLM]: Support text replacement offsets in the remaining old-format processors

src/transformers/models/gemma3/processing_gemma3.py

@@ -12,16 +12,12 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import re
 
 from ...feature_extraction_utils import BatchFeature
 from ...image_utils import ImageInput, make_nested_list_of_images
 from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
-from ...utils import auto_docstring, logging, to_py_obj
-
-
-logger = logging.get_logger(__name__)
+from ...utils import auto_docstring, to_py_obj
 
 
 class Gemma3ProcessorKwargs(ProcessingKwargs, total=False):
@@ -55,6 +51,7 @@ def __init__(
         self.image_seq_length = image_seq_length
         self.image_token_id = tokenizer.image_token_id
         self.boi_token = tokenizer.boi_token
+        self.image_token = tokenizer.boi_token
         image_tokens_expanded = "".join([tokenizer.image_token] * image_seq_length)
         self.full_image_sequence = f"\n\n{tokenizer.boi_token}{image_tokens_expanded}{tokenizer.eoi_token}\n\n"
 
@@ -72,67 +69,53 @@ def __call__(
         text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None,
         **kwargs: Unpack[Gemma3ProcessorKwargs],
     ) -> BatchFeature:
-        if text is None and images is None:
-            raise ValueError("Provide at least one of `text` or `images`.")
+        # Create empty text to be replaced with placeholders
+        if images is not None and not text:
+            images = self.image_processor.fetch_images(images)
+            text = [" ".join([self.boi_token] * len(images)) for images in make_nested_list_of_images(images)]
 
-        output_kwargs = self._merge_kwargs(
-            Gemma3ProcessorKwargs,
-            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
-            **kwargs,
-        )
+        model_inputs = super().__call__(images=images, text=text, **kwargs)
+        if "mm_token_type_ids" in model_inputs:
+            model_inputs["token_type_ids"] = model_inputs.pop("mm_token_type_ids")
+        return model_inputs
 
-        if isinstance(text, str):
-            text = [text]
-        elif not isinstance(text, list) and not isinstance(text[0], str):
-            raise TypeError("Invalid input text. Please provide a string, or a list of strings")
+    def validate_inputs(
+        self,
+        images: ImageInput | None = None,
+        text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None = None,
+        **kwargs: Unpack[ProcessingKwargs],
+    ):
+        super().validate_inputs(images=images, text=text, **kwargs)
 
-        image_inputs = {}
-        if images is not None:
-            images = self.image_processor.fetch_images(images)
+        if images is not None and text is not None:
             batched_images = make_nested_list_of_images(images)
-            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
-
-            # Create empty text to be replaced with placeholders
-            if not text:
-                text = [" ".join([self.boi_token] * len(images)) for images in batched_images]
-
             if len(batched_images) != len(text):
                 raise ValueError(
                     f"Received inconsistently sized batches of images ({len(batched_images)}) and text ({len(text)})."
                 )
 
-            # Replace image tokens by the full expanded sequence
-            num_crops = to_py_obj(image_inputs.pop("num_crops"))
-            batch_num_crops = [[num_crops.pop(0) for _ in range(len(images))] for images in batched_images]
-            for batch_idx, (prompt, images, num_crops) in enumerate(zip(text, batched_images, batch_num_crops)):
-                image_indexes = [m.start() for m in re.finditer(self.boi_token, prompt)]
-
-                if len(images) != len(image_indexes):
+            for prompt, images in zip(text, batched_images):
+                if len(images) != prompt.count(self.boi_token):
                     raise ValueError(
-                        f"Prompt contained {len(image_indexes)} image tokens but received {len(images)} images."
+                        f"Prompt contained {prompt.count(self.boi_token)} image tokens but received {len(images)} images."
                     )
 
-                # Insert additional image tokens for Pan-and-Scan crops
-                for num, idx in reversed(list(zip(num_crops, image_indexes))):
-                    if num:
-                        formatted_image_text = (
-                            f"Here is the original image {self.boi_token} and here are some crops to help you see better "
-                            + " ".join([self.boi_token] * num)
-                        )
-                        prompt = prompt[:idx] + formatted_image_text + prompt[idx + len(self.boi_token) :]
-                        text[batch_idx] = prompt
-
-            # Expand placeholder image tokens to the full image token sequence
-            text = [prompt.replace(self.boi_token, self.full_image_sequence) for prompt in text]
-
-        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
-        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", False)
-        text_inputs = self.tokenizer(text=text, **output_kwargs["text_kwargs"])
-        self._check_special_mm_tokens(text, text_inputs, modalities=["image"])
-
-        if return_mm_token_type_ids:
-            text_inputs["token_type_ids"] = self.create_mm_token_type_ids(text_inputs["input_ids"])
-        return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)
+    def _check_special_mm_tokens(self, text: list[str], text_inputs: "BatchFeature", modalities: list[str]):
+        """
+        Checks that number of special tokens in text and processed text is same. The count can be different
+        if tokenized text was truncated, leading to issues in model code.
+        """
+        token_str = self.tokenizer.image_token
+        token_id = self.tokenizer.image_token_id
+        if token_str is not None and token_id is not None:
+            ids_count = [list(ids).count(token_id) for ids in text_inputs["input_ids"]]
+            text_count = [sample.count(token_str) for sample in text]
+
+        if ids_count != text_count:
+            raise ValueError(
+                f"Mismatch in `image` token count between text and `input_ids`. Got ids={ids_count} and text={text_count}. "
+                "Likely due to `truncation='max_length'`. Please disable truncation or increase `max_length`."
+            )
 
     def _get_num_multimodal_tokens(self, image_sizes=None, **kwargs):
         """
@@ -165,13 +148,16 @@ def model_input_names(self) -> list[str]:
     def unused_input_names(self) -> list[str]:
         return ["num_crops"]
 
-    @property
-    def image_token(self) -> list[str]:
-        logger.warning_once(
-            "Deprecated: `processor.image_token` will switch from returning "
-            "`tokenizer.image_token` to `tokenizer.boi_token` in v5.11."
+    def replace_image_token(self, image_inputs: dict, image_idx: int, **kwargs) -> str:
+        num_crops = to_py_obj(image_inputs["num_crops"])[image_idx]
+        if not num_crops:
+            return self.full_image_sequence
+
+        # Insert additional image tokens for Pan-and-Scan crops
+        return (
+            f"Here is the original image {self.full_image_sequence} and here are some crops to help you see better "
+            + " ".join([self.full_image_sequence] * num_crops)
         )
-        return self.tokenizer.image_token
 
 
 __all__ = ["Gemma3Processor"]

src/transformers/models/granite_speech/processing_granite_speech.py

@@ -13,23 +13,28 @@
 # limitations under the License.
 """Processor class for Granite Speech."""
 
-from typing import Union
-
+from ...audio_utils import AudioInput
 from ...feature_extraction_utils import BatchFeature
-from ...processing_utils import ProcessorMixin
+from ...processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_python import PreTokenizedInput, TextInput
-from ...utils import auto_docstring, is_torch_available, logging
-from ...utils.import_utils import requires_backends
-
+from ...utils import auto_docstring
 
-if is_torch_available():
-    import torch
 
-logger = logging.get_logger(__name__)
+class GraniteSpeechProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": True,
+        },
+        "audio_kwargs": {
+            "device": "cpu",
+        },
+    }
 
 
 @auto_docstring
 class GraniteSpeechProcessor(ProcessorMixin):
+    valid_processor_kwargs = GraniteSpeechProcessorKwargs
+
     def __init__(
         self,
         audio_processor,
@@ -50,49 +55,13 @@ def __init__(
     def __call__(
         self,
         text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput],
-        audio: Union["torch.Tensor", list["torch.Tensor"]] = None,
+        audio: AudioInput | None = None,
         device: str = "cpu",
-        **kwargs,
+        **kwargs: Unpack[GraniteSpeechProcessorKwargs],
     ) -> BatchFeature:
-        requires_backends(self, ["torch"])
-
         text = self._get_validated_text(text)
-        prompt_strings = text
-
-        if audio is not None:
-            # NOTE - we intentionally avoid throwing for potentially misaligned
-            # text / audio inputs here because some inference engines will
-            # trigger the conditions due to the way they call multimodal
-            # processors, e.g., vLLM.
-            audio_inputs = self.audio_processor(audio, device=device)
-
-            # TODO (@alex-jw-brooks); we should add a util to get_num_audio_tokens
-            # from feature lengths and call it here, rather than returning it
-            # from the feature extractor.
-            audio_embed_sizes = audio_inputs.pop("audio_embed_sizes")
-
-            # Expand the audio placeholders to match the feature dims; this
-            # is similar to how many VLMs handle image tokens, e.g., llava next
-            prompt_strings = []
-            num_replaced = 0
-            for sample in text:
-                while self.audio_token in sample:
-                    sample = sample.replace(
-                        self.audio_token,
-                        "<placeholder>" * audio_embed_sizes[num_replaced],
-                        1,
-                    )
-                    num_replaced += 1
-                prompt_strings.append(sample)
-
-            prompt_strings = [sample.replace("<placeholder>", self.audio_token) for sample in prompt_strings]
-        else:
-            audio_inputs = {}
-
-        if "padding" not in kwargs:
-            kwargs["padding"] = True
-        text_inputs = self.tokenizer(prompt_strings, **kwargs)
-        return BatchFeature(data={**text_inputs, **audio_inputs})
+        kwargs.setdefault("audio_kwargs", {}).setdefault("device", device)
+        return super().__call__(text=text, audio=audio, **kwargs)
 
     def _get_validated_text(self, text: str | list) -> list[str]:
         if isinstance(text, str):
@@ -101,5 +70,23 @@ def _get_validated_text(self, text: str | list) -> list[str]:
             return text
         raise TypeError("Invalid text provided! Text should be a string or list of strings.")
 
+    def _process_audio(self, audio: AudioInput, **kwargs):
+        # Audio samples are already collated
+        if len(audio) == 1 and audio[0].ndim == 2:
+            audio = audio[0]
+        audio_inputs = self.audio_processor(audio, device=kwargs["device"])
+
+        audio_replacements = [self.replace_audio_token(audio_inputs, audio_idx=idx) for idx in range(len(audio))]
+        return audio_inputs, audio_replacements
+
+    def replace_audio_token(self, audio_inputs: dict, audio_idx: int, **kwargs) -> str:
+        num_audio_tokens = audio_inputs["audio_embed_sizes"][audio_idx]
+        return self.audio_token * num_audio_tokens
+
+    @property
+    def unused_input_names(self) -> list[str]:
+        "Input names returned always by subprocessors but not used in model's `forward`"
+        return ["audio_embed_sizes"]
+
 
 __all__ = ["GraniteSpeechProcessor"]