[internvl] fix chat template (#37656)

* fix chat template

* update

* update conversion

* rename `fake_image_token` in tests

Author: zucchini-nlp

docs/source/en/model_doc/internvl.md

@@ -257,6 +257,7 @@ InternVL models can also handle video inputs. Here is an example of how to perfo
 ...     add_generation_prompt=True,
 ...     tokenize=True,
 ...     return_dict=True,
+...     num_frames=8,
 >>> ).to(model.device, dtype=torch.float16)
 
 >>> output = model.generate(**inputs, max_new_tokens=25)

src/transformers/models/internvl/convert_internvl_weights_to_hf.py

@@ -312,6 +312,7 @@ def write_tokenizer(save_dir: str, push_to_hub: bool = False, path: str = None,
                 "start_image_token": "<img>",
                 "end_image_token": "</img>",
                 "context_image_token": "<IMG_CONTEXT>",
+                "video_token": "<video>",
             },
         )
         tokenizer.model_max_length = CONTEXT_LENGTH

src/transformers/models/internvl/processing_internvl.py

@@ -14,13 +14,11 @@
 # limitations under the License.
 
 
-from functools import partial
-from typing import Dict, List, Optional, Union
+from typing import List, Optional, Union
 
 import numpy as np
 
 from transformers.processing_utils import (
-    AllKwargsForChatTemplate,
     ImagesKwargs,
     ProcessingKwargs,
     ProcessorMixin,
@@ -34,6 +32,7 @@
     VideoInput,
     VideoMetadata,
     concatenate_list,
+    load_video,
     make_batched_videos,
     make_flat_list_of_images,
 )
@@ -75,20 +74,12 @@ class InternVLProcessor(ProcessorMixin):
             image_seq_length = (config.image_size // config.patch_size) ** 2 * (config.scale_factor**2)
         chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
             in a chat into a tokenizable string.
-        fake_image_token (`str`, *optional*, defaults to `"<image>"`):
-            The token to use for the image placeholder in the text. This token will be replaced by the
-            appropriate image tokens when processing the text with images.
-        fake_video_token (`str`, *optional*, defaults to `"<video>"`):
-            The token to use for the video placeholder in the text. This token will be replaced by the
-            appropriate image tokens when processing the text with videos.
     """
 
     attributes = ["image_processor", "tokenizer"]
     valid_kwargs = [
         "chat_template",
         "image_seq_length",
-        "fake_image_token",
-        "fake_video_token",
     ]
     image_processor_class = "AutoImageProcessor"
     tokenizer_class = "AutoTokenizer"
@@ -99,16 +90,14 @@ def __init__(
         tokenizer=None,
         image_seq_length: int = 256,
         chat_template=None,
-        fake_image_token="<image>",
-        fake_video_token="<video>",
         **kwargs,
     ):
         self.image_seq_length = image_seq_length
-        self.fake_image_token = fake_image_token
-        self.fake_video_token = fake_video_token
         self.start_image_token = tokenizer.start_image_token
         self.end_image_token = tokenizer.end_image_token
-        self.context_image_token = tokenizer.context_image_token
+        self.image_token = tokenizer.context_image_token
+        self.video_token = tokenizer.video_token
+        self.image_token_id = tokenizer.context_image_token_id
 
         super().__init__(image_processor, tokenizer, chat_template=chat_template, **kwargs)
 
@@ -131,24 +120,24 @@ def _insert_media_placeholders(
         video_index = 0
         processed_text = []
         image_video_patches = []
+        replace_strings = []
         # Support interleaved image and video in prompts:
         # Processed patches of images and videos are inserted in `image_video_patches` in the order they appear in the prompts
         for prompt in text:
             new_prompt = prompt
-            while self.fake_image_token in new_prompt or self.fake_video_token in new_prompt:
-                if self.fake_image_token in new_prompt and (
-                    self.fake_video_token not in new_prompt
-                    or new_prompt.index(self.fake_image_token) < new_prompt.index(self.fake_video_token)
+            while self.image_token in new_prompt or self.video_token in new_prompt:
+                if self.image_token in new_prompt and (
+                    self.video_token not in new_prompt
+                    or new_prompt.index(self.image_token) < new_prompt.index(self.video_token)
                 ):
                     # Get the slice of patches corresponding to the current image
                     start_index = image_num_patches_indices[image_index - 1] if image_index > 0 else 0
                     end_index = image_num_patches_indices[image_index]
                     image_video_patches.append(image_pixel_values[start_index:end_index])
                     # Replace the corresponding image placeholder with the correct number of image tokens
-                    new_prompt = new_prompt.replace(
-                        self.fake_image_token,
-                        f"{self.start_image_token}{self.context_image_token * self.image_seq_length * image_num_patches[image_index]}{self.end_image_token}",
-                        1,
+                    new_prompt = new_prompt.replace(self.image_token, "<placeholder>", 1)
+                    replace_strings.append(
+                        f"{self.start_image_token}{self.image_token * self.image_seq_length * image_num_patches[image_index]}{self.end_image_token}"
                     )
                     image_index += 1
                 else:
@@ -163,11 +152,15 @@ def _insert_media_placeholders(
                     # Get the number of patches per frame and replace the video placeholder with the correct number of image tokens
                     num_patches = list(video_num_patches[current_patch_index:end_patch_index])
                     video_prompt = "\n".join(
-                        f"Frame{i + 1}: {self.start_image_token}{self.context_image_token * self.image_seq_length * num_patches[i]}{self.end_image_token}"
+                        f"Frame{i + 1}: {self.start_image_token}{self.image_token * self.image_seq_length * num_patches[i]}{self.end_image_token}"
                         for i in range(len(num_patches))
                     )
-                    new_prompt = new_prompt.replace(self.fake_video_token, video_prompt, 1)
+                    replace_strings.append(video_prompt)
+                    new_prompt = new_prompt.replace(self.video_token, "<placeholder>", 1)
                     video_index += 1
+            while "<placeholder>" in new_prompt:
+                replace_str = replace_strings.pop(0)
+                new_prompt = new_prompt.replace("<placeholder>", replace_str, 1)
             processed_text.append(new_prompt)
 
         return processed_text, image_video_patches, image_index, video_index
@@ -269,9 +262,11 @@ def __call__(
             # Concatenate the interleaved image and video patches (function agnostic to the patches type (list, numpy array, torch tensor))
             image_videos_inputs = {"pixel_values": concatenate_list(image_video_patches)}
 
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
         text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        self._check_special_mm_tokens(text, text_inputs, modalities=["image"])
 
-        return BatchFeature(data={**text_inputs, **image_videos_inputs})
+        return BatchFeature(data={**text_inputs, **image_videos_inputs}, tensor_type=return_tensors)
 
     def sample_indices_fn(
         self, metadata: VideoMetadata, num_frames: int = None, initial_shift: Union[bool, float, int] = True
@@ -290,15 +285,13 @@ def sample_indices_fn(
         Returns:
             `np.ndarray`: Array of frame indices to sample.
         """
+        num_frames = num_frames if num_frames is not None else metadata.total_num_frames
+
         if initial_shift is True:
             initial_shift = metadata.total_num_frames / num_frames / 2
-        if num_frames is not None:
-            indices = np.arange(
-                initial_shift, metadata.total_num_frames, metadata.total_num_frames / num_frames
-            ).astype(int)
-        else:
-            indices = np.arange(initial_shift, metadata.total_num_frames).astype(int)
-
+        indices = np.arange(initial_shift, metadata.total_num_frames, metadata.total_num_frames / num_frames).astype(
+            int
+        )
         return indices
 
     def batch_decode(self, *args, **kwargs):
@@ -321,58 +314,39 @@ def model_input_names(self):
         image_processor_input_names = self.image_processor.model_input_names
         return list(tokenizer_input_names) + list(image_processor_input_names)
 
-    # Add model-specific video sampling method when applying the template
-    def apply_chat_template(
+    # TODO: raushan, has to be public method under `VideoProcessorBase` when API is added
+    def _load_video_for_model(
         self,
-        conversation: Union[List[Dict[str, str]], List[List[Dict[str, str]]]],
-        chat_template: Optional[str] = None,
-        num_frames: int = 8,
-        initial_shift: Union[bool, float, int] = True,
-        video_load_backend="pyav",
-        **kwargs: Unpack[AllKwargsForChatTemplate],
-    ):
+        video: Union[str, "VideoInput"],
+        num_frames: Optional[int],
+        backend: str = "pyav",
+        initial_shift: bool = True,
+        **kwargs,
+    ) -> np.array:
         """
-        Similar to the `apply_chat_template` method on tokenizers, this method applies a Jinja template to input
-        conversations to turn them into a single tokenizable string.
-
-        The input is expected to be in the following format, where each message content is a list consisting of text and
-        optionally image or video inputs. One can also provide an image, video, URL or local path which will be used to form
-        `pixel_values` when `return_dict=True`. If not provided, one will get only the formatted text, optionally tokenized text.
-
-        conversation = [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "image", "image": "https://www.ilankelman.org/stopsigns/australia.jpg"},
-                    {"type": "text", "text": "Please describe this image in detail."},
-                ],
-            },
-        ]
+        Loads `video` to a numpy array.
 
         Args:
-            conversation (`Union[List[Dict, [str, str]], List[List[Dict[str, str]]]]`):
-                The conversation to format.
-            chat_template (`Optional[str]`, *optional*):
-                The Jinja template to use for formatting the conversation. If not provided, the tokenizer's
-                chat template is used.
-            num_frames (`int`, *optional*, defaults to 8):
-                Number of frames to sample from a video when using the default `sample_indices_fn`.
-            initial_shift (`bool`, `float` or `int`, defaults to `0`):
-                The initial shift to apply when sampling frames using the default `sample_indices_fn`.
-                If `True`, the shift is set so that frames are sampled from the middle of the video.
+            video (`str` or `VideoInput`):
+                The video to convert to the numpy array format. Can be a link to video or local path.
+            num_frames (`int`, *optional*):
+                Number of frames to sample uniformly. If not passed, the whole video is loaded.
+            backend (`str`, *optional*, defaults to `"pyav"`):
+                The backend to use when loading the video. Can be any of ["decord", "pyav", "opencv", "torchvision"]. Defaults to "pyav".
+            initial_shift (`bool`, *optional*, defaults to `True`):
+                The initial shift to apply when sampling frames. If `True`, the shift is set so that frames are sampled from the middle of the video.
+
+        Returns:
+            Tuple[`np.array`, Dict]: A tuple containing:
+                - Numpy array of frames in RGB (shape: [num_frames, height, width, 3]).
+                - Metadata dictionary.
         """
-        sample_indices_fn = kwargs.pop(
-            "sample_indices_fn", partial(self.sample_indices_fn, num_frames=num_frames, initial_shift=initial_shift)
-        )
 
-        return super().apply_chat_template(
-            conversation,
-            chat_template,
-            video_load_backend=video_load_backend,
-            num_frames=num_frames,
-            sample_indices_fn=sample_indices_fn,
-            **kwargs,
-        )
+        def sample_indices_fn_func(metadata, **fn_kwargs):
+            return self.sample_indices_fn(metadata, num_frames=num_frames, initial_shift=initial_shift, **fn_kwargs)
+
+        video, metadata = load_video(video, backend=backend, sample_indices_fn=sample_indices_fn_func)
+        return video, metadata
 
 
 __all__ = ["InternVLProcessor"]

tests/models/internvl/test_modeling_internvl.py

@@ -296,7 +296,9 @@ def test_qwen2_small_model_integration_generate(self):
         url = "http://images.cocodataset.org/val2017/000000039769.jpg"
         image = Image.open(requests.get(url, stream=True).raw)
 
-        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        prompt = (
+            "<|im_start|>user\n<IMG_CONTEXT>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        )
         inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
         with torch.no_grad():
             generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
@@ -314,7 +316,9 @@ def test_qwen2_small_model_integration_forward(self):
         url = "http://images.cocodataset.org/val2017/000000039769.jpg"
         image = Image.open(requests.get(url, stream=True).raw)
 
-        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        prompt = (
+            "<|im_start|>user\n<IMG_CONTEXT>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        )
         inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
 
         # Forward
@@ -378,8 +382,8 @@ def test_qwen2_small_model_integration_batched_generate(self):
         )
         # Prepare inputs
         prompt = [
-            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
-            "<|im_start|>user\n<image>\nDescribe this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT>\nDescribe this image<|im_end|>\n<|im_start|>assistant\n",
         ]
         image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
         image2 = Image.open(requests.get("https://www.ilankelman.org/stopsigns/australia.jpg", stream=True).raw)
@@ -414,8 +418,8 @@ def test_qwen2_small_model_integration_batched_generate_multi_image(self):
         )
         # Prepare inputs
         prompt = [
-            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
-            "<|im_start|>user\n<image><image>\nWhat are the differences between these two images?<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT><IMG_CONTEXT>\nWhat are the differences between these two images?<|im_end|>\n<|im_start|>assistant\n",
         ]
         image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
         image2 = Image.open(
@@ -485,6 +489,7 @@ def test_qwen2_medium_model_integration_video(self):
             tokenize=True,
             return_dict=True,
             return_tensors="pt",
+            num_frames=8,
         ).to(torch_device, dtype=torch.float16)
 
         output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
@@ -552,6 +557,7 @@ def test_qwen2_small_model_integration_interleaved_images_videos(self):
             return_dict=True,
             return_tensors="pt",
             padding=True,
+            num_frames=8,
         ).to(torch_device, dtype=torch.bfloat16)
 
         output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
@@ -601,7 +607,9 @@ def test_llama_small_model_integration_generate(self):
         url = "http://images.cocodataset.org/val2017/000000039769.jpg"
         image = Image.open(requests.get(url, stream=True).raw)
 
-        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        prompt = (
+            "<|im_start|>user\n<IMG_CONTEXT>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        )
         inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
         with torch.no_grad():
             generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
@@ -619,7 +627,9 @@ def test_llama_small_model_integration_forward(self):
         url = "http://images.cocodataset.org/val2017/000000039769.jpg"
         image = Image.open(requests.get(url, stream=True).raw)
 
-        prompt = "<|im_start|>user\n<image>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        prompt = (
+            "<|im_start|>user\n<IMG_CONTEXT>\nPlease describe the image explicitly.<|im_end|>\n<|im_start|>assistant\n"
+        )
         inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
 
         # Forward
@@ -687,8 +697,8 @@ def test_llama_small_model_integration_batched_generate(self):
         )
         # Prepare inputs
         prompt = [
-            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
-            "<|im_start|>user\n<image>\nDescribe this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT>\nDescribe this image<|im_end|>\n<|im_start|>assistant\n",
         ]
         image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
         image2 = Image.open(requests.get("https://www.ilankelman.org/stopsigns/australia.jpg", stream=True).raw)
@@ -724,8 +734,8 @@ def test_llama_small_model_integration_batched_generate_multi_image(self):
         )
         # Prepare inputs
         prompt = [
-            "<|im_start|>user\n<image>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
-            "<|im_start|>user\n<image><image>\nWhat are the difference between these two images?<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT>\nWrite a haiku for this image<|im_end|>\n<|im_start|>assistant\n",
+            "<|im_start|>user\n<IMG_CONTEXT><IMG_CONTEXT>\nWhat are the difference between these two images?<|im_end|>\n<|im_start|>assistant\n",
         ]
         image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
         image2 = Image.open(
@@ -795,6 +805,7 @@ def test_llama_medium_model_integration_video(self):
             tokenize=True,
             return_dict=True,
             return_tensors="pt",
+            num_frames=8,
         ).to(torch_device, dtype=torch.float16)
 
         output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
@@ -862,6 +873,7 @@ def test_llama_small_model_integration_interleaved_images_videos(self):
             return_dict=True,
             return_tensors="pt",
             padding=True,
+            num_frames=8,
         ).to(torch_device, dtype=torch.bfloat16)
 
         output = model.generate(**inputs, do_sample=False, max_new_tokens=25)