VLMs: patch_size -> num_image_tokens in processing

docs/source/en/model_doc/blip-2.md

@@ -40,6 +40,10 @@ The original code can be found [here](https://github.com/salesforce/LAVIS/tree/5
 - BLIP-2 can be used for conditional text generation given an image and an optional text prompt. At inference time, it's recommended to use the [`generate`] method.
 - One can use [`Blip2Processor`] to prepare images for the model, and decode the predicted tokens ID's back to text.
 
+> [!NOTE]
+> BLIP models after release v4.46 will raise warnings about adding `processor.num_query_tokens = {{num_query_tokens}}` and expand model embeddings layer to add special `<image>` token. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you. Adding these attributes means that BLIP will add the number of query tokens required per image and expand the text with as many `<image>` placeholders as there will be query tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there wil be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.num_query_tokens` and model embeddings expansion can be done by following [this link](https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042).
+
 ## Resources
 
 A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with BLIP-2.

docs/source/en/model_doc/instructblip.md

@@ -33,6 +33,10 @@ The original code can be found [here](https://github.com/salesforce/LAVIS/tree/m
 
 InstructBLIP uses the same architecture as [BLIP-2](blip2) with a tiny but important difference: it also feeds the text prompt (instruction) to the Q-Former.
 
+> [!NOTE]
+> BLIP models after release v4.46 will raise warnings about adding `processor.num_query_tokens = {{num_query_tokens}}` and expand model embeddings layer to add special `<image>` token. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you. Adding these attributes means that BLIP will add the number of query tokens required per image and expand the text with as many `<image>` placeholders as there will be query tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there wil be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.num_query_tokens` and model embeddings expansion can be done by following [this link](https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042).
+
 ## InstructBlipConfig
 
 [[autodoc]] InstructBlipConfig

docs/source/en/model_doc/instructblipvideo.md

@@ -35,6 +35,10 @@ The original code can be found [here](https://github.com/salesforce/LAVIS/tree/m
 
 - The model was trained by sampling 4 frames per video, so it's recommended to sample 4 frames
 
+> [!NOTE]
+> BLIP models after release v4.46 will raise warnings about adding `processor.num_query_tokens = {{num_query_tokens}}` and expand model embeddings layer to add special `<image>` token. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you. Adding these attributes means that BLIP will add the number of query tokens required per image and expand the text with as many `<image>` placeholders as there will be query tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there wil be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.num_query_tokens` and model embeddings expansion can be done by following [this link](https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042).
+
 ## InstructBlipVideoConfig
 
 [[autodoc]] InstructBlipVideoConfig

docs/source/en/model_doc/llava.md

@@ -40,6 +40,13 @@ The original code can be found [here](https://github.com/haotian-liu/LLaVA/tree/
 
 - Note the model has not been explicitly trained to process multiple images in the same prompt, although this is technically possible, you may experience inaccurate results.
 
+
+> [!NOTE]
+> LLaVA models after release v4.46 will raise warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you.
+Adding these attributes means that LLaVA will try to infer the number of image tokens required per image and expand the text with as many `<image>` placeholders as there will be tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there will be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`. The `num_additional_image_tokens` should be `1` if the vision backbone adds a CLS token or `0` if nothing extra is added to the vision patches.
+
+
 ### Single image inference
 
 For best results, we recommend users to use the processor's `apply_chat_template()` method to format your prompt correctly. For that you need to construct a conversation history, passing in a plain string will not format your prompt. Each message in the conversation history for chat templates is a dictionary with keys "role" and "content". The "content" should be a list of dictionaries, for "text" and "image" modalities, as follows:

docs/source/en/model_doc/llava_next.md

@@ -53,6 +53,12 @@ The original code can be found [here](https://github.com/haotian-liu/LLaVA/tree/
 </Tip>
 
 
+> [!NOTE]
+> LLaVA models after release v4.46 will raise warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you.
+Adding these attributes means that LLaVA will try to infer the number of image tokens required per image and expand the text with as many `<image>` placeholders as there will be tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there will be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`. The `num_additional_image_tokens` should be `1` if the vision backbone adds a CLS token or `0` if nothing extra is added to the vision patches.
+
+
 - Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. You can use the processor's `apply_chat_template` to format your prompts correctly. For that you have to construct a conversation history, passing a plain string will not format your prompt. Each message in the conversation history for chat templates is a dictionary with keys "role" and "content". The "content" should be a list of dictionaries, for "text" and "image" modalities. Below is an example of how to do that and the list of formats accepted by each checkpoint.
 
 We will use [llava-v1.6-mistral-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-mistral-7b-hf) and a conversation history of text and image. Each content field has to be a list of dicts, as follows:

docs/source/en/model_doc/llava_next_video.md

@@ -50,6 +50,12 @@ The original code can be found [here](https://github.com/LLaVA-VL/LLaVA-NeXT/tre
 </Tip>
 
 
+> [!NOTE]
+> LLaVA models after release v4.46 will raise warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you.
+Adding these attributes means that LLaVA will try to infer the number of image tokens required per image and expand the text with as many `<image>` placeholders as there will be tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there will be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`. The `num_additional_image_tokens` should be `1` if the vision backbone adds a CLS token or `0` if nothing extra is added to the vision patches.
+
+
 - Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. You can use tokenizer's `apply_chat_template` to format your prompts correctly. Below is an example of how to do that.
 
 We will use [LLaVA-NeXT-Video-7B-hf](https://huggingface.co/llava-hf/LLaVA-NeXT-Video-7B-hf) and a conversation history of videos and images. Each content field has to be a list of dicts, as follows:

docs/source/en/model_doc/video_llava.md

@@ -54,6 +54,12 @@ This model was contributed by [RaushanTurganbay](https://huggingface.co/RaushanT
 The original code can be found [here](https://github.com/PKU-YuanGroup/Video-LLaVA).
 
 
+> [!NOTE]
+> LLaVA models after release v4.46 will raise warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you.
+Adding these attributes means that LLaVA will try to infer the number of image tokens required per image and expand the text with as many `<image>` placeholders as there will be tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there will be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`. The `num_additional_image_tokens` should be `1` if the vision backbone adds a CLS token or `0` if nothing extra is added to the vision patches.
+
+
 ## Usage example
 
 ### Single Media Mode

docs/source/en/model_doc/vipllava.md

@@ -39,6 +39,12 @@ This model was contributed by [Younes Belkada](https://huggingface.co/ybelkada)
 
 - Note the model has not been explicitly trained to process multiple images in the same prompt, although this is technically possible, you may experience inaccurate results.
 
+> [!NOTE]
+> LLaVA models after release v4.46 will raise warnings about adding `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. It is strongly recommended to add the attributes to the processor if you own the model checkpoint, or open a PR if it is not owned by you.
+Adding these attributes means that LLaVA will try to infer the number of image tokens required per image and expand the text with as many `<image>` placeholders as there will be tokens. Usually it is around 500 tokens per image, so make sure that the text is not truncated as otherwise there will be failure when merging the embeddings.
+The attributes can be obtained from model config, as `model.config.vision_config.patch_size` or `model.config.vision_feature_select_strategy`. The `num_additional_image_tokens` should be `1` if the vision backbone adds a CLS token or `0` if nothing extra is added to the vision patches.
+
+
 - For better results, we recommend users to use the processor's `apply_chat_template()` method to format your prompt correctly. For that you need to construct a conversation history, passing in a plain string will not format your prompt. Each message in the conversation history for chat templates is a dictionary with keys "role" and "content". The "content" should be a list of dictionaries, for "text" and "image" modalities, as follows:
 
 ```python

src/transformers/models/llava/processing_llava.py

@@ -51,6 +51,9 @@ class LlavaProcessor(ProcessorMixin):
             The tokenizer is a required input.
         patch_size (`int`, *optional*):
             Patch size from the vision tower.
+        num_additional_image_tokens (`int`, *optional*, defaults to 0):
+            Number of additional tokens added to the image embeddings, such as CLS (+1). If the backbone has no CLS or other
+            extra tokens appended, no need to set this arg.
         vision_feature_select_strategy (`str`, *optional*):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Shoudl be same as in model's config
@@ -61,7 +64,13 @@ class LlavaProcessor(ProcessorMixin):
     """
 
     attributes = ["image_processor", "tokenizer"]
-    valid_kwargs = ["chat_template", "patch_size", "vision_feature_select_strategy", "image_token"]
+    valid_kwargs = [
+        "chat_template",
+        "patch_size",
+        "vision_feature_select_strategy",
+        "image_token",
+        "num_additional_image_tokens",
+    ]
     image_processor_class = "AutoImageProcessor"
     tokenizer_class = "AutoTokenizer"
 
@@ -70,12 +79,14 @@ def __init__(
         image_processor=None,
         tokenizer=None,
         patch_size=None,
+        num_additional_image_tokens=0,
         vision_feature_select_strategy=None,
         chat_template=None,
         image_token="<image>",  # set the default and let users change if they have peculiar special tokens in rare cases
         **kwargs,
     ):
         self.patch_size = patch_size
+        self.num_additional_image_tokens = num_additional_image_tokens
         self.vision_feature_select_strategy = vision_feature_select_strategy
         self.image_token = image_token
         super().__init__(image_processor, tokenizer, chat_template=chat_template)
@@ -147,9 +158,11 @@ def __call__(
                 # Replace the image token with the expanded image token sequence
                 pixel_values = image_inputs["pixel_values"]
                 height, width = get_image_size(to_numpy_array(pixel_values[0]))
-                num_image_tokens = (height // self.patch_size) * (width // self.patch_size) + 1
+                num_image_tokens = (height // self.patch_size) * (
+                    width // self.patch_size
+                ) + self.num_additional_image_tokens
                 if self.vision_feature_select_strategy == "default":
-                    num_image_tokens -= 1
+                    num_image_tokens -= self.num_additional_image_tokens
 
                 prompt_strings = []
                 for sample in text:
@@ -158,8 +171,9 @@ def __call__(
             else:
                 logger.warning_once(
                     "Expanding inputs for image tokens in LLaVa should be done in processing. "
-                    "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
-                    "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
+                    "Please add `patch_size`, `num_additional_image_tokens` and `vision_feature_select_strategy` to the model's processing config or set directly "
+                    "with `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` "
+                    "and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
                     "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
                 )
 

src/transformers/models/llava_next/processing_llava_next.py

@@ -54,6 +54,9 @@ class LlavaNextProcessor(ProcessorMixin):
             The tokenizer is a required input.
         patch_size (`int`, *optional*):
             Patch size from the vision tower.
+        num_additional_image_tokens (`int`, *optional*, defaults to 0):
+            Number of additional tokens added to the image embeddings, such as CLS (+1). If the backbone has no CLS or other
+            extra tokens appended, no need to set this arg.
         vision_feature_select_strategy (`str`, *optional*):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Shoudl be same as in model's config
@@ -64,7 +67,13 @@ class LlavaNextProcessor(ProcessorMixin):
     """
 
     attributes = ["image_processor", "tokenizer"]
-    valid_kwargs = ["chat_template", "patch_size", "vision_feature_select_strategy", "image_token"]
+    valid_kwargs = [
+        "chat_template",
+        "patch_size",
+        "vision_feature_select_strategy",
+        "image_token",
+        "num_additional_image_tokens",
+    ]
     image_processor_class = "AutoImageProcessor"
     tokenizer_class = "AutoTokenizer"
 
@@ -73,12 +82,14 @@ def __init__(
         image_processor=None,
         tokenizer=None,
         patch_size=None,
+        num_additional_image_tokens=0,
         vision_feature_select_strategy=None,
         chat_template=None,
         image_token="<image>",  # set the default and let users change if they have peculiar special tokens in rare cases
         **kwargs,
     ):
         self.patch_size = patch_size
+        self.num_additional_image_tokens = num_additional_image_tokens
         self.vision_feature_select_strategy = vision_feature_select_strategy
         self.image_token = image_token
         super().__init__(image_processor, tokenizer, chat_template=chat_template)
@@ -141,8 +152,9 @@ def __call__(
             if self.patch_size is None or self.vision_feature_select_strategy is None:
                 logger.warning_once(
                     "Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
-                    "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
-                    "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
+                    "Please add `patch_size`, `num_additional_image_tokens` and `vision_feature_select_strategy` to the model's processing config or set directly "
+                    "with `processor.patch_size = {{patch_size}}`, `processor.num_additional_image_tokens = {{num_additional_image_tokens}}` "
+                    "and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
                     "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
                 )
             else:
@@ -155,7 +167,7 @@ def __call__(
                         orig_height, orig_width = image_size
                         num_image_tokens = self._get_number_of_features(orig_height, orig_width, height, width)
                         if self.vision_feature_select_strategy == "default":
-                            num_image_tokens -= 1
+                            num_image_tokens -= self.num_additional_image_tokens
                         sample = sample.replace(self.image_token, "<placeholder>" * num_image_tokens, 1)
                     prompt_strings.append(sample)
                 prompt_strings = [sample.replace("<placeholder>", self.image_token) for sample in prompt_strings]
@@ -178,7 +190,7 @@ def _get_number_of_features(self, orig_height: int, orig_width: int, height: int
             orig_height, orig_width, patches_height, patches_width, scale_height, scale_width
         )
         # The base patch covers the entire image (+1 for the CLS)
-        base_features = patches_height * patches_width + 1
+        base_features = patches_height * patches_width + self.num_additional_image_tokens
         num_image_tokens = unpadded_features + newline_features + base_features
         return num_image_tokens