o ۷iD @sdZddlmZmZddlZddlmZddlm Z m Z ddl m Z m Z mZmZddlmZmZGd d d e d d ZGd dde d d ZdeededeeefddZdeeeedeeedededejf ddZdedededefddZGdd d eZd gZdS)!zProcessor class for Mllama.)OptionalUnionN) BatchFeature) ImageInputmake_nested_list_of_images) ImagesKwargsProcessingKwargsProcessorMixinUnpack)PreTokenizedInput TextInputc@seZdZUeeed<dS)MllamaImagesKwargsmax_image_tilesN)__name__ __module__ __qualname__rint__annotations__rrb/home/ubuntu/vllm_env/lib/python3.10/site-packages/transformers/models/mllama/processing_mllama.pyrs rF)totalc@s"eZdZUeed<dddiiZdS)MllamaProcessorKwargs images_kwargs image_kwargsrN)rrrrr _defaultsrrrrr s r input_idsimage_token_idreturncsfddt|D}t|dkrgSt|dkr |ddggSddt|dd|ddD}||dt|g|dd}|dddD]}|d|ddkrZ||d<|d}qJ|S)a Generate a cross-attention token mask for image tokens in the input sequence. This function identifies the positions of image tokens in the input sequence and creates a mask that defines which subsequent tokens each image token should attend to. Args: input_ids (list[int]): A list of token ids representing the input sequence. image_token_id (int): The id of the token used to represent images in the sequence. Returns: list[list[int]]: A list of [start, end] pairs, where each pair represents the range of tokens an image token should attend to. Notes: - If no image tokens are present, an empty list is returned. - For a single image token, it attends to all subsequent tokens until the end of the sequence. - For multiple image tokens, each attends to tokens up to the next image token or the end of the sequence. - Consecutive image tokens are treated as a group and attend to all subsequent tokens together. csg|] \}}|kr|qSrr).0itokenrrr @z2get_cross_attention_token_mask..rcSsg|]\}}||gqSrr)r loc1loc2rrrr$IN) enumeratelenzipappend)rrimage_token_locations vision_masks last_mask_end vision_maskrr#rget_cross_attention_token_mask*s  $  r3cross_attention_token_mask num_tiles max_num_tileslengthc Cst|}tdd|D}tj||||ftjd}tt||D]5\}\}} tt|| D]'\} \} } t| dkrT| \} }t||}|dkrH|}d||| || d| f<q-q |S)a Convert the cross attention mask indices to a cross attention mask 4D array. This function takes a sparse representation of cross attention masks and converts it to a dense 4D numpy array. The sparse representation is a nested list structure that defines attention ranges for each image in each batch item. Args: cross_attention_token_mask (list[list[list[int]]]): A nested list structure where: - The outer list represents the batch dimension. - The middle list represents different images within each batch item. - The inner list contains pairs of integers [start, end] representing token ranges for each image. num_tiles (list[list[int]]): A nested list structure specifying the number of tiles for each image in each batch item. max_num_tiles (int): The maximum possible number of tiles. length (int): The total sequence length of the input. Returns: np.ndarray: A 4D numpy array of shape (batch_size, length, max_num_images, max_num_tiles) The array contains `1` where attention is allowed and `0` where it is not. Note: - Special handling is done for cases where the end token is -1, which is interpreted as attending to the end of the sequence. cs|]}t|VqdSNr,)r masksrrr xz?convert_sparse_cross_attention_mask_to_dense..)shapedtyper'r&N)r,maxnpzerosint64r+r-min)r4r5r6r7 batch_sizemax_num_imagescross_attention_mask sample_idx sample_maskssample_num_tilesmask_idx locationsmask_num_tilesstartendrrr,convert_sparse_cross_attention_mask_to_denseZs"   rQprompt bos_token image_tokencCsP||vr|Sd}||r|t|d}|d7}||s ||||S)a\ Builds a string from the input prompt by adding `bos_token` if not already present. Args: prompt (`str`): The input prompt string. bos_token (`str`): The beginning of sentence token to be added. image_token (`str`): The image token used to identify the start of an image sequence. Returns: str: The modified prompt string with the `bos_token` added if necessary. Examples: >>> build_string_from_input("Hello world", "", "<|image|>") 'Hello world' >>> build_string_from_input("<|image|>Hello world", "", "<|image|>") '<|image|>Hello world' >>> build_string_from_input("Hello world", "", "<|image|>") 'Hello world' rNr&) startswithr,)rRrSrTnum_image_tokens_on_startrrrbuild_string_from_inputs  rWc seZdZdZddgZdZdZdfdd Z    dd ee d ee e e e e e e fd eed efd dZ dddZeddZZS)MllamaProcessora Constructs a Mllama processor which wraps [`MllamaImageProcessor`] and [`PretrainedTokenizerFast`] into a single processor that inherits both the image processor and tokenizer functionalities. See the [`~MllamaProcessor.__call__`] and [`~OwlViTProcessor.decode`] for more information. The preferred way of passing kwargs is as a dictionary per modality, see usage example below. ```python from transformers import MllamaProcessor from PIL import Image processor = MllamaProcessor.from_pretrained("meta-llama/Llama-3.2-11B-Vision") processor( images=your_pil_image, text=["<|image|>If I had to write a haiku for this one"], images_kwargs = {"size": {"height": 448, "width": 448}}, text_kwargs = {"padding": "right"}, common_kwargs = {"return_tensors": "pt"}, ) ``` Args: image_processor ([`MllamaImageProcessor`]): The image processor is a required input. tokenizer ([`PreTrainedTokenizer`, `PreTrainedTokenizerFast`]): 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. image_processor tokenizerMllamaImageProcessorPreTrainedTokenizerFastNcsbt|dsd|_||j|_n|j|_|j|_d|_||j|_|j|_tj|||ddS)NrTz <|image|>z<|python_tag|>) chat_template) hasattrrTconvert_tokens_to_idsr python_tokenpython_token_idrSsuper__init__)selfrYrZr] __class__rrrcs zMllamaProcessor.__init__imagestextkwargsrc s|dur |dur tdjtfdjji|}|d}d|d<|d}|d} i} |durt|tr8|g}nt|ttfrHt dd |DsLtd fd d |D} fd d |D}| dd} j|fi|} j || dgdfdd | dD}| | dg}|durj |}t|}dd |D}|durtdd | Drt dd | Dstdt| dkr|| ks||kr|durtdd}t|t| kr|| krd}n||krd}td| d|d||dur j |fi|}| d}| ||dur6|dur6fdd | dD}t||j jtd d | dDd!}|| d"<| dd}t| |d#}|S)$a& Main method to prepare text(s) and image(s) to be fed as input to the model. This method forwards the `text` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] if `text` is not `None` to encode the text. To prepare the image(s), this method forwards the `images` arguments to MllamaImageProcessor's [`~MllamaImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring 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. Both channels-first and channels-last formats are supported. 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. Returned when `text` is not `None`. - **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`. TODO: add aspect_ratio_ids and aspect_ratio_mask and cross_attention_mask Nz'You must specify either text or images.tokenizer_init_kwargs text_kwargsreturn_tensorsr common_kwargscss|]}t|tVqdSr9) isinstancestrr trrrr<sz+MllamaProcessor.__call__..zAInvalid input text. Please provide a string, or a list of stringscg|]}|jqSr)countrTrprdrrr$r*z,MllamaProcessor.__call__..csg|] }t|jjqSr)rWrSrT)r text_itemrtrrr$r% padding_sideimage) modalitiescrrr)rsrr token_idsrtrrr$ r*rrcSsg|]}t|qSrr:)r samplerrrr$'scss|]}|dkVqdS)rNr)r batch_imgrrrr<*r=zaIf a batch of text is provided, there should be either no images or at least one image per samplez@No image were provided, but there are image tokens in the promptzZMake sure to pass your images as a nested list, where each sub-list holds images per batchzhIf you activated truncation with `max_length`, increase the `max_length` so image tokens aren't cropped.z)The number of image tokens in each text (zA) should be the same as the number of provided images per batch (z). r5csg|]}t|jqSr)r3rryrtrrr$Hs csr8r9r:)r rrrrr<Or=)r5r6r7rH)data tensor_type) ValueError _merge_kwargsrrZ init_kwargsrnrolisttupleallpop_check_special_mm_tokensupdaterY fetch_imagesranysumrQrrAr)rdrgrhaudiovideosri output_kwargsrkrrmr~n_images_in_text_encodingn_images_in_idsn_images_in_images add_messageimage_featuresr5r4rHrl batch_featurerrtr__call__s'            zMllamaProcessor.__call__TFcKs|jj|f||d|S)a Post-process the output of the model to decode the text. Args: generated_outputs (`torch.Tensor` or `np.ndarray`): The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)` or `(sequence_length,)`. skip_special_tokens (`bool`, *optional*, defaults to `True`): Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method. clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`): Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method. **kwargs: Additional arguments to be passed to the tokenizer's `batch_decode method`. Returns: `list[str]`: The decoded text. )skip_special_tokensclean_up_tokenization_spaces)rZ batch_decode)rdgenerated_outputsrrrirrrpost_process_image_text_to_textXsz/MllamaProcessor.post_process_image_text_to_textcCs0|jj}|jj}dd|D}t||dgS)NcSsg|]}|dkr|qS)r5r)r namerrrr$zr*z5MllamaProcessor.model_input_names..rH)rZmodel_input_namesrYr)rdtokenizer_input_namesimage_processor_input_namesrrrrssz!MllamaProcessor.model_input_namesr9)NNNN)TF)rrr__doc__ attributesimage_processor_classtokenizer_classrcrrrr r rr rrrrpropertyr __classcell__rrrerrXs. y rX)rtypingrrnumpyrBfeature_extraction_utilsr image_utilsrrprocessing_utilsrr r r tokenization_utils_baser r rrrrr3ndarrayrQrorWrX__all__rrrrs2 " 0  0% P