# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/colqwen2/modular_colqwen2.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_colqwen2.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # coding=utf-8 # Copyright 2025 The HuggingFace Inc. team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # 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. from typing import Optional, Union from ...feature_extraction_utils import BatchFeature from ...image_utils import ImageInput, is_valid_image from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack from ...tokenization_utils_base import PreTokenizedInput, TextInput from ...utils import is_torch_available if is_torch_available(): import torch class ColQwen2ProcessorKwargs(ProcessingKwargs, total=False): _defaults = { "text_kwargs": { "padding": "longest", }, "images_kwargs": { "data_format": "channels_first", "do_convert_rgb": True, }, "common_kwargs": {"return_tensors": "pt"}, } class ColQwen2Processor(ProcessorMixin): r""" Constructs a ColQwen2 processor which wraps a Qwen2VLProcessor and special methods to process images and queries, as well as to compute the late-interaction retrieval score. [`ColQwen2Processor`] offers all the functionalities of [`Qwen2VLProcessor`]. See the [`~Qwen2VLProcessor.__call__`] for more information. Args: image_processor ([`Qwen2VLImageProcessor`], *optional*): The image processor is a required input. tokenizer ([`Qwen2TokenizerFast`], *optional*): The tokenizer is a required input. chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string. visual_prompt_prefix (`str`, *optional*): A string that gets tokenized and prepended to the image tokens. query_prefix (`str`, *optional*): A prefix to be used for the query. """ attributes = ["image_processor", "tokenizer"] image_processor_class = "AutoImageProcessor" tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast") def __init__( self, image_processor=None, tokenizer=None, chat_template=None, visual_prompt_prefix: Optional[str] = None, query_prefix: Optional[str] = None, **kwargs, ): super().__init__(image_processor, tokenizer, chat_template=chat_template) self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token if visual_prompt_prefix is None: visual_prompt_prefix = "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe the image.<|im_end|><|endoftext|>" self.visual_prompt_prefix = visual_prompt_prefix if query_prefix is None: query_prefix = "Query: " self.query_prefix = query_prefix def __call__( self, images: ImageInput = None, text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None, audio=None, videos=None, **kwargs: Unpack[ColQwen2ProcessorKwargs], ) -> BatchFeature: """ Main method to prepare for the model either (1) one or several texts, either (2) one or several image(s). This method is a custom wrapper around the Qwen2VLProcessor's [`~Qwen2VLProcessor.__call__`] method adapted for the ColQwen2 model. It cannot process both text and images at the same time. When preparing the the text(s), this method forwards the `text` and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`]. When preparing the the image(s), this method forwards the `images` and `kwargs` arguments to Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`]. Please refer to the doctsring of the above two methods for more information. Args: images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`): The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a number of channels, H and W are image height and width. text (`str`, `list[str]`, `list[list[str]]`): The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set `is_split_into_words=True` (to lift the ambiguity with a batch of sequences). return_tensors (`str` or [`~utils.TensorType`], *optional*): If set, will return tensors of a particular framework. Acceptable values are: - `'tf'`: Return TensorFlow `tf.constant` objects. - `'pt'`: Return PyTorch `torch.Tensor` objects. - `'np'`: Return NumPy `np.ndarray` objects. - `'jax'`: Return JAX `jnp.ndarray` objects. Returns: [`BatchFeature`]: A [`BatchFeature`] with the following fields: - **input_ids** -- List of token ids to be fed to a model. - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not `None`). - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`. """ output_kwargs = self._merge_kwargs( ColQwen2ProcessorKwargs, tokenizer_init_kwargs=self.tokenizer.init_kwargs, **kwargs, ) suffix = output_kwargs["text_kwargs"].pop("suffix", None) return_token_type_ids = True if suffix is not None else False if text is None and images is None: raise ValueError("Either text or images must be provided") if text is not None and images is not None: raise ValueError("Only one of text or images can be processed at a time") if images is not None: if is_valid_image(images): images = [images] elif isinstance(images, list) and is_valid_image(images[0]): pass elif not (isinstance(images, list) and isinstance(images[0], list) and is_valid_image(images[0][0])): raise ValueError("images must be an image, list of images or list of list of images") texts_doc = [self.visual_prompt_prefix] * len(images) image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"]) image_grid_thw = image_inputs["image_grid_thw"] if image_grid_thw is not None: merge_length = self.image_processor.merge_size**2 index = 0 for i in range(len(texts_doc)): while self.image_token in texts_doc[i]: texts_doc[i] = texts_doc[i].replace( self.image_token, "<|placeholder|>" * (image_grid_thw[index].prod() // merge_length), 1 ) index += 1 texts_doc[i] = texts_doc[i].replace("<|placeholder|>", self.image_token) text_inputs = self.tokenizer( texts_doc, return_token_type_ids=False, **output_kwargs["text_kwargs"], ) return_data = BatchFeature(data={**text_inputs, **image_inputs}) # NOTE: The following adjustment ensures correct behavior with DDP on multiple GPUs. offsets = return_data["image_grid_thw"][:, 1] * return_data["image_grid_thw"][:, 2] # (batch_size,) # Split the pixel_values tensor into a list of tensors, one per image pixel_values = list( torch.split(return_data["pixel_values"], offsets.tolist()) ) # [(num_patches_image_0, pixel_values), ..., (num_patches_image_n, pixel_values)] # Pad the list of pixel_value tensors to the same length along the sequence dimension return_data["pixel_values"] = torch.nn.utils.rnn.pad_sequence( pixel_values, batch_first=True ) # (batch_size, max_num_patches, pixel_values) if return_token_type_ids: labels = return_data["input_ids"].masked_fill(return_data["token_type_ids"] == 0, -100) return_data.update({"labels": labels}) return return_data elif text is not None: if isinstance(text, str): text = [text] elif not (isinstance(text, list) and isinstance(text[0], str)): raise ValueError("Text must be a string or a list of strings") if suffix is None: suffix = self.query_augmentation_token * 10 texts_query: list[str] = [] for query in text: augmented_query = self.query_prefix + query + suffix texts_query.append(augmented_query) batch_query = self.tokenizer( texts_query, return_token_type_ids=False, **output_kwargs["text_kwargs"], ) return batch_query def _get_num_multimodal_tokens(self, image_sizes=None, **kwargs): """ Computes the number of placeholder tokens needed for multimodal inputs with the given sizes. Args: image_sizes (list[list[str]], *optional*): The input sizes formatted as (height, width) per each image. Returns: dict[str, list[int]]: A dictionary mapping each modality ("image", "video", "audio") to a list containing the number of placeholder tokens required. If the model doesn't accept a certain modality or no input sizes are provided, the dict value is set to an empty list. """ vision_data = {} if image_sizes is not None: num_image_tokens = [self.image_seq_length] * len(image_sizes) num_image_patches = [1] * len(image_sizes) vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches}) return MultiModalData(**vision_data) def batch_decode(self, *args, **kwargs): """ This method forwards all its arguments to GemmaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please refer to the docstring of this method for more information. """ return self.tokenizer.batch_decode(*args, **kwargs) def decode(self, *args, **kwargs): """ This method forwards all its arguments to GemmaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to the docstring of this method for more information. """ return self.tokenizer.decode(*args, **kwargs) @property def model_input_names(self): tokenizer_input_names = self.tokenizer.model_input_names image_processor_input_names = self.image_processor.model_input_names return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names)) @property def query_augmentation_token(self) -> str: """ Return the query augmentation token. Query augmentation buffers are used as reasoning buffers during inference. """ return self.tokenizer.pad_token def process_images( self, images: ImageInput = None, **kwargs: Unpack[ColQwen2ProcessorKwargs], ) -> BatchFeature: """ Prepare for the model one or several image(s). This method is a wrapper around the `__call__` method of the ColQwen2Processor's [`ColQwen2Processor.__call__`]. This method forwards the `images` and `kwargs` arguments to the image processor. Args: images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`): The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a number of channels, H and W are image height and width. return_tensors (`str` or [`~utils.TensorType`], *optional*): If set, will return tensors of a particular framework. Acceptable values are: - `'tf'`: Return TensorFlow `tf.constant` objects. - `'pt'`: Return PyTorch `torch.Tensor` objects. - `'np'`: Return NumPy `np.ndarray` objects. - `'jax'`: Return JAX `jnp.ndarray` objects. Returns: [`BatchFeature`]: A [`BatchFeature`] with the following fields: - **input_ids** -- List of token ids to be fed to a model. - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not `None`). - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`. """ return self.__call__(images=images, **kwargs) def process_queries( self, text: Union[TextInput, list[TextInput]], **kwargs: Unpack[ColQwen2ProcessorKwargs], ) -> BatchFeature: """ Prepare for the model one or several texts. This method is a wrapper around the `__call__` method of the ColQwen2Processor's [`ColQwen2Processor.__call__`]. This method forwards the `text` and `kwargs` arguments to the tokenizer. Args: text (`str`, `list[str]`, `list[list[str]]`): The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set `is_split_into_words=True` (to lift the ambiguity with a batch of sequences). return_tensors (`str` or [`~utils.TensorType`], *optional*): If set, will return tensors of a particular framework. Acceptable values are: - `'tf'`: Return TensorFlow `tf.constant` objects. - `'pt'`: Return PyTorch `torch.Tensor` objects. - `'np'`: Return NumPy `np.ndarray` objects. - `'jax'`: Return JAX `jnp.ndarray` objects. Returns: [`BatchFeature`]: A [`BatchFeature`] with the following fields: - **input_ids** -- List of token ids to be fed to a model. - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not `None`). """ return self.__call__(text=text, **kwargs) def score_retrieval( self, query_embeddings: Union["torch.Tensor", list["torch.Tensor"]], passage_embeddings: Union["torch.Tensor", list["torch.Tensor"]], batch_size: int = 128, output_dtype: Optional["torch.dtype"] = None, output_device: Union["torch.device", str] = "cpu", ) -> "torch.Tensor": """ Compute the late-interaction/MaxSim score (ColBERT-like) for the given multi-vector query embeddings (`qs`) and passage embeddings (`ps`). For ColQwen2, a passage is the image of a document page. Because the embedding tensors are multi-vector and can thus have different shapes, they should be fed as: (1) a list of tensors, where the i-th tensor is of shape (sequence_length_i, embedding_dim) (2) a single tensor of shape (n_passages, max_sequence_length, embedding_dim) -> usually obtained by padding the list of tensors. Args: query_embeddings (`Union[torch.Tensor, list[torch.Tensor]`): Query embeddings. passage_embeddings (`Union[torch.Tensor, list[torch.Tensor]`): Passage embeddings. batch_size (`int`, *optional*, defaults to 128): Batch size for computing scores. output_dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): The dtype of the output tensor. If `None`, the dtype of the input embeddings is used. output_device (`torch.device` or `str`, *optional*, defaults to "cpu"): The device of the output tensor. Returns: `torch.Tensor`: A tensor of shape `(n_queries, n_passages)` containing the scores. The score tensor is saved on the "cpu" device. """ if len(query_embeddings) == 0: raise ValueError("No queries provided") if len(passage_embeddings) == 0: raise ValueError("No passages provided") if query_embeddings[0].device != passage_embeddings[0].device: raise ValueError("Queries and passages must be on the same device") if query_embeddings[0].dtype != passage_embeddings[0].dtype: raise ValueError("Queries and passages must have the same dtype") if output_dtype is None: output_dtype = query_embeddings[0].dtype scores: list[torch.Tensor] = [] for i in range(0, len(query_embeddings), batch_size): batch_scores: list[torch.Tensor] = [] batch_queries = torch.nn.utils.rnn.pad_sequence( query_embeddings[i : i + batch_size], batch_first=True, padding_value=0 ) for j in range(0, len(passage_embeddings), batch_size): batch_passages = torch.nn.utils.rnn.pad_sequence( passage_embeddings[j : j + batch_size], batch_first=True, padding_value=0 ) batch_scores.append( torch.einsum("bnd,csd->bcns", batch_queries, batch_passages).max(dim=3)[0].sum(dim=2) ) scores.append(torch.cat(batch_scores, dim=1).to(output_dtype).to(output_device)) return torch.cat(scores, dim=0) __all__ = ["ColQwen2Processor"]