# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/glm_image/modular_glm_image.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_glm_image.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2025 the HuggingFace Team. All rights reserved. # # 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. import math import torch from ...feature_extraction_utils import BatchFeature from ...image_utils import ImageInput from ...processing_utils import ImagesKwargs, ProcessingKwargs, ProcessorMixin, Unpack from ...tokenization_utils_base import PreTokenizedInput, TextInput from ...utils.import_utils import requires class GlmImageImagesKwargs(ImagesKwargs, total=False): """ target_h (`int`): Height of the target image to be generated. target_w (`int`): Width of the target image to be generated. """ target_h: int target_w: int class GlmImageProcessorKwargs(ProcessingKwargs, total=False): _defaults = { "text_kwargs": { "padding": False, "return_mm_token_type_ids": False, }, "images_kwargs": { "target_h": 1152, "target_w": 768, }, } images_kwargs: GlmImageImagesKwargs @requires(backends=("torch",)) class GlmImageProcessor(ProcessorMixin): r""" Constructs a GLM-Image processor which wraps a GLM-Image image processor and a GLM-Image tokenizer into a single processor. [`~GlmImageProcessor.__call__`] and [`~GlmImageProcessor.decode`] for more information. Args: image_processor ([`GlmImageProcessor`], *optional*): The image processor is a required input. tokenizer ([`PreTrainedTokenizerFast`], *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. """ valid_processor_kwargs = GlmImageProcessorKwargs model_input_names = ["input_ids", "attention_mask", "pixel_values", "image_grid_thw", "images_per_sample"] def __init__(self, image_processor=None, tokenizer=None, chat_template=None, **kwargs): self.image_token = tokenizer.image_token self.grid_bos_token = tokenizer.grid_bos_token self.grid_eos_token = tokenizer.grid_eos_token self.bos_token = tokenizer.bos_token self.image_token_id = tokenizer.convert_tokens_to_ids(self.image_token) super().__init__(image_processor, tokenizer, chat_template=chat_template) def __call__( self, images: ImageInput | None = None, text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None, **kwargs: Unpack[GlmImageProcessorKwargs], ) -> BatchFeature: """ Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text` and `kwargs` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] if `text` is not `None` to encode the text. 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: - `'pt'`: Return PyTorch `torch.Tensor` objects. - `'np'`: Return NumPy `np.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`. - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`. """ output_kwargs = self._merge_kwargs( GlmImageProcessorKwargs, tokenizer_init_kwargs=self.tokenizer.init_kwargs, **kwargs, ) model_inputs = super().__call__(images=images, text=text, **output_kwargs) if text is None: # early exit if cond only on text return model_inputs target_h = output_kwargs["images_kwargs"].get("target_h") target_w = output_kwargs["images_kwargs"].get("target_w") return_tensors = output_kwargs["text_kwargs"].get("return_tensors") is_text_to_image = images is None # Count images per sample by counting image tokens in each text batch_size = len(text) if not isinstance(text, str) else 1 images_per_sample = [] for i in range(batch_size): images_per_sample.append(text[i].count(self.image_token)) # Build prompt with target shape and combine grids in a single loop # Format: [sample0_source_grids..., sample0_target_grids, sample1_source_grids..., sample1_target_grids, ...] # Note: In image2image mode, batches are homogeneous (same number of source images per sample) num_source_images = images_per_sample[0] if images_per_sample else 0 all_grids = [] for i in range(batch_size): token_h, token_w, prev_h, prev_w = self._get_target_shape(height=target_h, width=target_w) # Add source grids for this sample (i2i mode only) if not is_text_to_image and num_source_images > 0: start_idx = i * num_source_images all_grids.append(model_inputs["image_grid_thw"][start_idx : start_idx + num_source_images]) # Add target grid for this sample all_grids.append( self._build_target_image_grid_thw( token_h=token_h, token_w=token_w, prev_token_h=prev_h, prev_token_w=prev_w, is_text_to_image=is_text_to_image, ) ) model_inputs["image_grid_thw"] = torch.cat(all_grids, dim=0) # Store images_per_sample for later use (add target images count) # Each sample will have: source_images + target_images (typically 2 for t2i, 1 for i2i) num_target_grids = 2 if is_text_to_image else 1 model_inputs["images_per_sample"] = torch.tensor( [num_source_images + num_target_grids] * batch_size, dtype=torch.long ) return BatchFeature(data=model_inputs, tensor_type=return_tensors) def prepare_inputs_layout( self, images: ImageInput | None = None, text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None = None, **kwargs: Unpack[GlmImageImagesKwargs], ): images, text, *_ = super().prepare_inputs_layout(images=images, text=text, **kwargs) processed_text = text if text is not None: processed_text = [] target_h = kwargs["images_kwargs"].get("target_h", None) target_w = kwargs["images_kwargs"].get("target_w", None) for sample in text: token_h, token_w, prev_token_h, prev_token_w = self._get_target_shape(height=target_h, width=target_w) if images is None: sample = f"{sample}{self.grid_bos_token}{token_h} {token_w}{self.grid_eos_token}{self.grid_bos_token}{prev_token_h} {prev_token_w}{self.grid_eos_token}{self.bos_token}" else: sample = f"{sample}{self.grid_bos_token}{token_h} {token_w}{self.grid_eos_token}{self.bos_token}" processed_text.append(sample) return images, processed_text, None, None def validate_inputs( self, images: ImageInput | None = None, text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None = None, **kwargs: Unpack[GlmImageImagesKwargs], ): super().validate_inputs(images=images, text=text) if text is None and images is None: raise ValueError("You must provide at least one of `text` or `images`.") # Validate homogeneity for i2i mode images_per_sample = [] for i in range(len(text)): images_per_sample.append(text[i].count(self.image_token)) if images is not None and images_per_sample and len(set(images_per_sample)) != 1: raise ValueError( f"In image-to-image mode, all samples must have the same number of source images. " f"Got different counts: {images_per_sample}" ) def _process_images(self, images: ImageInput, **kwargs): # Kwargs used only in processor, if we pass them down to image processor it raises an error kwargs.pop("target_h", None) kwargs.pop("target_w", None) return super()._process_images(images, **kwargs) def replace_image_token(self, image_inputs: dict, image_idx: int, **kwargs) -> str: merge_length = self.image_processor.merge_size**2 num_image_tokens = image_inputs["image_grid_thw"][image_idx].prod() // merge_length return self.image_token * num_image_tokens def _get_target_shape( self, height: int, width: int, ) -> tuple[str, int, int, int, int]: factor = 32 height = (height // factor) * factor width = (width // factor) * factor token_h = height // factor token_w = width // factor ratio = token_h / token_w prev_token_h = int(math.sqrt(ratio) * (factor // 2)) prev_token_w = int(math.sqrt(1 / ratio) * (factor // 2)) return token_h, token_w, prev_token_h, prev_token_w @staticmethod def _build_target_image_grid_thw( token_h: int, token_w: int, prev_token_h: int, prev_token_w: int, is_text_to_image: bool = True, ): if is_text_to_image: # Text-to-image: 2 target grids (large + small preview) return torch.tensor( [ [1, token_h, token_w], [1, prev_token_h, prev_token_w], ], ) else: # Image-to-image: 1 target grid only return torch.tensor( [ [1, token_h, token_w], ], ) __all__ = ["GlmImageProcessor"]