Granite Vision Support (#35579)

* Add multimodal granite support

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

Support multiple image feature layres

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Remove failing validation for visual encoders with no cls

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Update llava based models / configs to support list of feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Add tests for multiple feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use conditional instead of except for misaligned feature shapes

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

* crop cls from each hidden state

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

* Fix formatting

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Support single vision feature int in vipllava

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix typo in vision feature selection strategy validation

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

* Add tentative integration test for granite vision models

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

* Add granite vision docs

Replace multimodal granite refs with granite vision

Add granite vision / llava next alias

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

* Use image url in granitevision example

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

---------

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

docs/source/en/_toctree.yml

@@ -448,6 +448,8 @@
         title: Granite
       - local: model_doc/granitemoe
         title: GraniteMoe
+      - local: model_doc/granitevision
+        title: GraniteVision
       - local: model_doc/helium
         title: Helium
       - local: model_doc/herbert

docs/source/en/model_doc/granitevision.md

@@ -0,0 +1,90 @@
+<!--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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Granite Vision
+
+## Overview
+
+The Granite Vision model is a variant of [LLaVA-NeXT](llava_next), leveraging a [Granite](granite) language model alongside a [SigLIP](SigLIP) visual encoder. It utilizes multiple concatenated vision hidden states as its image features, similar to [VipLlava](vipllava). It also uses a larger set of image grid pinpoints than the original LlaVa-NeXT models to support additional aspect ratios.
+
+Tips:
+- This model is loaded into Transformers as an instance of LlaVA-Next. The usage and tips from [LLaVA-NeXT](llava_next) apply to this model as well.
+
+- You can apply the chat template on the tokenizer / processor in the same way as well. Example chat format:
+```bash
+"<|user|>\nWhat’s shown in this image?\n<|assistant|>\nThis image shows a red stop sign.<|end_of_text|><|user|>\nDescribe the image in more details.\n<|assistant|>\n"
+```
+
+Sample inference:
+```python
+from transformers import LlavaNextProcessor, LlavaNextForConditionalGeneration
+from PIL import Image
+import requests
+
+# Note: These docs were written prior to the public model release,
+# and this path is subject to change.
+# Please see https://huggingface.co/ibm-granite for the current model list.
+model_path = "ibm-granite/granite-3.1-2b-instruct-vision"
+processor = LlavaNextProcessor.from_pretrained(model_path)
+
+model = LlavaNextForConditionalGeneration.from_pretrained(model_path).to("cuda")
+
+# prepare image and text prompt, using the appropriate prompt template
+url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
+
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image", "url": url},
+            {"type": "text", "text": "What is shown in this image?"},
+        ],
+    },
+]
+inputs = processor.apply_chat_template(
+    conversation,
+    add_generation_prompt=True,
+    tokenize=True,
+    return_dict=True,
+    return_tensors="pt"
+).to("cuda")
+
+
+# autoregressively complete prompt
+output = model.generate(**inputs, max_new_tokens=100)
+
+print(processor.decode(output[0], skip_special_tokens=True))
+```
+
+This model was contributed by [Alexander Brooks](https://huggingface.co/abrooks9944).
+
+## LlavaNextConfig
+
+[[autodoc]] LlavaNextConfig
+
+## LlavaNextImageProcessor
+
+[[autodoc]] LlavaNextImageProcessor
+    - preprocess
+
+## LlavaNextProcessor
+
+[[autodoc]] LlavaNextProcessor
+
+## LlavaNextForConditionalGeneration
+
+[[autodoc]] LlavaNextForConditionalGeneration
+    - forward

src/transformers/models/auto/configuration_auto.py

@@ -134,6 +134,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("gptsan-japanese", "GPTSanJapaneseConfig"),
         ("granite", "GraniteConfig"),
         ("granitemoe", "GraniteMoeConfig"),
+        ("granitevision", "LlavaNextConfig"),
         ("graphormer", "GraphormerConfig"),
         ("grounding-dino", "GroundingDinoConfig"),
         ("groupvit", "GroupViTConfig"),
@@ -458,6 +459,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("gptsan-japanese", "GPTSAN-japanese"),
         ("granite", "Granite"),
         ("granitemoe", "GraniteMoeMoe"),
+        ("granitevision", "LLaVA-NeXT"),
         ("graphormer", "Graphormer"),
         ("grounding-dino", "Grounding DINO"),
         ("groupvit", "GroupViT"),
@@ -729,6 +731,7 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict(
         ("siglip_vision_model", "siglip"),
         ("chinese_clip_vision_model", "chinese_clip"),
         ("rt_detr_resnet", "rt_detr"),
+        ("granitevision", "llava_next"),
     ]
 )
 

src/transformers/models/llava/configuration_llava.py

@@ -46,8 +46,10 @@ class LlavaConfig(PretrainedConfig):
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]]`, *optional*, defaults to -2):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         image_seq_length (`int`, *optional*, defaults to 576):
             Sequence length of one image embedding.
         multimodal_projector_bias (`bool`, *optional*, defaults to `True`):

src/transformers/models/llava/modeling_llava.py

@@ -87,8 +87,12 @@ class LlavaCausalLMOutputWithPast(ModelOutput):
 class LlavaMultiModalProjector(nn.Module):
     def __init__(self, config: LlavaConfig):
         super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
         self.linear_1 = nn.Linear(
-            config.vision_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
         )
         self.act = ACT2FN[config.projector_hidden_act]
         self.linear_2 = nn.Linear(
@@ -208,8 +212,10 @@ LLAVA_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
             is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
             model's internal embedding lookup matrix.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`.
@@ -270,7 +276,10 @@ class LlavaForConditionalGeneration(LlavaPreTrainedModel, GenerationMixin):
         return self.language_model.get_decoder()
 
     def get_image_features(
-        self, pixel_values: torch.FloatTensor, vision_feature_layer: int, vision_feature_select_strategy: str
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+        vision_feature_select_strategy: str,
     ):
         """
         Obtains image last hidden states from the vision tower and apply multimodal projection.
@@ -278,23 +287,35 @@ class LlavaForConditionalGeneration(LlavaPreTrainedModel, GenerationMixin):
         Args:
             pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
                The tensors corresponding to the input images.
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
         Returns:
             image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
         """
-        image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
+        if vision_feature_select_strategy not in ["default", "full"]:
-        # this is not memory efficient at all (output_hidden_states=True) will save all the hidden stated.
-        selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
-        if vision_feature_select_strategy == "default":
-            selected_image_feature = selected_image_feature[:, 1:]
-        elif vision_feature_select_strategy == "full":
-            selected_image_feature = selected_image_feature
-        else:
             raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}")
+
+        # this is not memory efficient at all (output_hidden_states=True) will save all the hidden states.
+        image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
+            if vision_feature_select_strategy == "default":
+                selected_image_feature = selected_image_feature[:, 1:]
+        else:
+            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            # For default; crop CLS from each hidden state in the hidden state pool
+            if vision_feature_select_strategy == "default":
+                hs_pool = [hs[:, 1:] for hs in hs_pool]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
         image_features = self.multi_modal_projector(selected_image_feature)
         return image_features
 
@@ -392,7 +413,7 @@ class LlavaForConditionalGeneration(LlavaPreTrainedModel, GenerationMixin):
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layer: Optional[int] = None,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
         vision_feature_select_strategy: Optional[str] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,

src/transformers/models/llava_next/configuration_llava_next.py

@@ -46,8 +46,10 @@ class LlavaNextConfig(PretrainedConfig):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
             If `"full"`, the full vision features are used.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]]`, *optional*, defaults to -2):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         image_grid_pinpoints (`List`, *optional*, defaults to `[[336, 672], [672, 336], [672, 672], [1008, 336], [336, 1008]]`):
             A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
             of the form `(height, width)`.

src/transformers/models/llava_next/modeling_llava_next.py

@@ -195,8 +195,12 @@ class LlavaNextCausalLMOutputWithPast(ModelOutput):
 class LlavaNextMultiModalProjector(nn.Module):
     def __init__(self, config: LlavaNextConfig):
         super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
         self.linear_1 = nn.Linear(
-            config.vision_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
         )
         self.act = ACT2FN[config.projector_hidden_act]
         self.linear_2 = nn.Linear(
@@ -319,8 +323,10 @@ LLAVA_NEXT_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
             is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
             model's internal embedding lookup matrix.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
@@ -664,18 +670,22 @@ class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel, GenerationMixi
                 image_feature = image_feature[1:]
                 height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
 
-                if vision_feature_select_strategy == "default":
-                    expected_num_patches = height * width
-                elif vision_feature_select_strategy == "full":
-                    expected_num_patches = height * width + 1
-                if expected_num_patches != base_image_feature.shape[0]:
-                    raise ValueError("The number of patches is not consistent with the image size.")
-
                 num_patch_height, num_patch_width = get_anyres_image_grid_shape(
                     image_sizes[image_idx],
                     self.config.image_grid_pinpoints,
                     self.config.vision_config.image_size,
                 )
+
+                if (
+                    np.prod(image_feature.shape) % (num_patch_height * num_patch_width * height * width) != 0
+                    and vision_feature_select_strategy == "default"
+                ):
+                    logger.warning_once(
+                        "Image feature shape does not line up with the provided patch size. "
+                        "You may be using the `default` vision_feature_select_strategy with a"
+                        " visual encoder that does not have CLS."
+                    )
+
                 image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                 image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                 image_feature = image_feature.flatten(1, 2).flatten(2, 3)
@@ -706,7 +716,7 @@ class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel, GenerationMixi
         self,
         pixel_values: torch.FloatTensor,
         image_sizes: torch.Tensor,
-        vision_feature_layer: int,
+        vision_feature_layer: Union[int, List[int]],
         vision_feature_select_strategy: str,
     ):
         """
@@ -717,8 +727,10 @@ class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel, GenerationMixi
                The tensors corresponding to the input images.
             image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                 Actual image size of each images (H, W).
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -744,11 +756,19 @@ class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel, GenerationMixi
             raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
 
         image_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
         if vision_feature_select_strategy == "default":
             selected_image_feature = selected_image_feature[:, 1:]
         elif vision_feature_select_strategy == "full":
             selected_image_feature = selected_image_feature
+
         image_features = self.multi_modal_projector(selected_image_feature)
         image_features = torch.split(image_features, image_num_patches, dim=0)
         return image_features
@@ -765,7 +785,7 @@ class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel, GenerationMixi
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layer: Optional[int] = None,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
         vision_feature_select_strategy: Optional[str] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,

src/transformers/models/llava_next_video/configuration_llava_next_video.py

@@ -50,8 +50,10 @@ class LlavaNextVideoConfig(PretrainedConfig):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
             If `"full"`, the full vision features are used.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]]`, *optional*, defaults to -2):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         image_grid_pinpoints (`List`, *optional*, defaults to `[[336, 672], [672, 336], [672, 672], [1008, 336], [336, 1008]]`):
             A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
             of the form `(height, width)`.

src/transformers/models/llava_next_video/modeling_llava_next_video.py

@@ -180,8 +180,12 @@ class LlavaNextVideoPreTrainedModel(PreTrainedModel):
 class LlavaNextVideoMultiModalProjector(nn.Module):
     def __init__(self, config: LlavaNextVideoConfig):
         super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
         self.linear_1 = nn.Linear(
-            config.vision_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
         )
         self.act = ACT2FN[config.projector_hidden_act]
         self.linear_2 = nn.Linear(
@@ -356,8 +360,10 @@ LLAVA_NEXT_VIDEO_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
             is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
             model's internal embedding lookup matrix.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
@@ -699,18 +705,22 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
                 image_feature = image_feature[1:]
                 height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
 
-                if vision_feature_select_strategy == "default":
-                    expected_num_patches = height * width
-                elif vision_feature_select_strategy == "full":
-                    expected_num_patches = height * width + 1
-                if expected_num_patches != base_image_feature.shape[0]:
-                    raise ValueError("The number of patches is not consistent with the image size.")
-
                 num_patch_height, num_patch_width = get_anyres_image_grid_shape(
                     image_sizes[image_idx],
                     self.config.image_grid_pinpoints,
                     self.config.vision_config.image_size,
                 )
+
+                if (
+                    np.prod(image_feature.shape) % (num_patch_height * num_patch_width * height * width) != 0
+                    and vision_feature_select_strategy == "default"
+                ):
+                    logger.warning_once(
+                        "Image feature shape does not line up with the provided patch size. "
+                        "You may be using the `default` vision_feature_select_strategy with a"
+                        " visual encoder that does not have CLS."
+                    )
+
                 image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                 image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                 image_feature = image_feature.flatten(1, 2).flatten(2, 3)
@@ -741,7 +751,7 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         self,
         pixel_values: torch.FloatTensor,
         image_sizes: torch.Tensor,
-        vision_feature_layer: int,
+        vision_feature_layer: Union[int, List[int]],
         vision_feature_select_strategy: str,
     ):
         """
@@ -752,8 +762,10 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
                The tensors corresponding to the input images.
             image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                 Actual image size of each images (H, W).
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -779,7 +791,14 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
             raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
 
         image_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
         if vision_feature_select_strategy == "default":
             selected_image_feature = selected_image_feature[:, 1:]
         elif vision_feature_select_strategy == "full":
@@ -801,7 +820,7 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layer: Optional[int] = None,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
         vision_feature_select_strategy: Optional[str] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
@@ -1043,7 +1062,10 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         return model_inputs
 
     def get_video_features(
-        self, pixel_values: torch.FloatTensor, vision_feature_layer: int, vision_feature_select_strategy: str
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+        vision_feature_select_strategy: str,
     ):
         """
         Obtains video last hidden states from the vision tower and apply multimodal projection.
@@ -1051,8 +1073,10 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         Args:
             pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
                The tensors corresponding to the input video.
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -1063,7 +1087,15 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, Gene
         batch_size, frames, channels, height, width = pixel_values.shape
         pixel_values = pixel_values.reshape(batch_size * frames, channels, height, width)
         video_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_video_features = video_features.hidden_states[vision_feature_layer]
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_video_features = video_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [video_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_video_features = torch.cat(hs_pool, dim=-1)
+
         if vision_feature_select_strategy == "default":
             selected_video_features = selected_video_features[:, 1:]
         elif vision_feature_select_strategy == "full":

src/transformers/models/llava_next_video/modular_llava_next_video.py

@@ -64,8 +64,10 @@ class LlavaNextVideoConfig(PretrainedConfig):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
             If `"full"`, the full vision features are used.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]]`, *optional*, defaults to -2):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         image_grid_pinpoints (`List`, *optional*, defaults to `[[336, 672], [672, 336], [672, 672], [1008, 336], [336, 1008]]`):
             A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
             of the form `(height, width)`.
@@ -237,7 +239,7 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
         self,
         pixel_values: torch.FloatTensor,
         image_sizes: torch.Tensor,
-        vision_feature_layer: int,
+        vision_feature_layer: Union[int, List[int]],
         vision_feature_select_strategy: str,
     ):
         """
@@ -248,8 +250,10 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
                The tensors corresponding to the input images.
             image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                 Actual image size of each images (H, W).
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -275,7 +279,14 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
             raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
 
         image_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
         if vision_feature_select_strategy == "default":
             selected_image_feature = selected_image_feature[:, 1:]
         elif vision_feature_select_strategy == "full":
@@ -285,7 +296,10 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
         return image_features
 
     def get_video_features(
-        self, pixel_values: torch.FloatTensor, vision_feature_layer: int, vision_feature_select_strategy: str
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+        vision_feature_select_strategy: str,
     ):
         """
         Obtains video last hidden states from the vision tower and apply multimodal projection.
@@ -293,8 +307,10 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
         Args:
             pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
                The tensors corresponding to the input video.
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -305,7 +321,15 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
         batch_size, frames, channels, height, width = pixel_values.shape
         pixel_values = pixel_values.reshape(batch_size * frames, channels, height, width)
         video_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_video_features = video_features.hidden_states[vision_feature_layer]
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_video_features = video_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [video_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_video_features = torch.cat(hs_pool, dim=-1)
+
         if vision_feature_select_strategy == "default":
             selected_video_features = selected_video_features[:, 1:]
         elif vision_feature_select_strategy == "full":
@@ -327,7 +351,7 @@ class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layer: Optional[int] = None,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
         vision_feature_select_strategy: Optional[str] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,

src/transformers/models/llava_onevision/configuration_llava_onevision.py

@@ -49,8 +49,10 @@ class LlavaOnevisionConfig(PretrainedConfig):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
             If `"full"`, the full vision features are used.
-        vision_feature_layer (`int`, *optional*, defaults to -1):
+        vision_feature_layer (`Union[int, List[int]]`, *optional*, defaults to -1):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
             Aspect ratio used when processong image features. The default value is "anyres_max_9".
         image_grid_pinpoints (`List`, *optional*):

src/transformers/models/llava_onevision/modeling_llava_onevision.py

@@ -202,8 +202,12 @@ class LlavaOnevisionCausalLMOutputWithPast(ModelOutput):
 class LlavaOnevisionMultiModalProjector(nn.Module):
     def __init__(self, config: LlavaOnevisionConfig):
         super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
         self.linear_1 = nn.Linear(
-            config.vision_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
         )
         self.act = ACT2FN[config.projector_hidden_act]
         self.linear_2 = nn.Linear(
@@ -334,8 +338,10 @@ LLAVA_ONEVISION_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
             is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
             model's internal embedding lookup matrix.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
@@ -488,7 +494,7 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
         self,
         pixel_values: torch.FloatTensor,
         image_sizes: torch.Tensor,
-        vision_feature_layer: int,
+        vision_feature_layer: Union[int, List[int]],
         vision_feature_select_strategy: str,
     ):
         """
@@ -499,8 +505,10 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
                The tensors corresponding to the input images.
             image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                 Actual image size of each images (H, W).
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -526,7 +534,14 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
             raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
 
         image_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
         if vision_feature_select_strategy == "default":
             selected_image_feature = selected_image_feature[:, 1:]
         elif vision_feature_select_strategy == "full":
@@ -536,7 +551,10 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
         return image_features
 
     def get_video_features(
-        self, pixel_values: torch.FloatTensor, vision_feature_layer: int, vision_feature_select_strategy: str
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+        vision_feature_select_strategy: str,
     ):
         """
         Obtains video last hidden states from the vision tower, apply multimodal projection and pooling.
@@ -544,8 +562,10 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
         Args:
             pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
                The tensors corresponding to the input video.
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
@@ -556,7 +576,14 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
         batch_size, frames, channels, height, width = pixel_values.shape
         pixel_values = pixel_values.view(batch_size * frames, channels, height, width)
         video_features = self.vision_tower(pixel_values, output_hidden_states=True)
-        selected_video_feature = video_features.hidden_states[vision_feature_layer]
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_video_feature = video_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [video_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_video_feature = torch.cat(hs_pool, dim=-1)
 
         if vision_feature_select_strategy == "default":
             selected_video_feature = selected_video_feature[:, 1:]
@@ -582,7 +609,7 @@ class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, Gene
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layer: Optional[int] = None,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
         vision_feature_select_strategy: Optional[str] = None,
         vision_aspect_ratio: Optional[str] = None,
         labels: Optional[torch.LongTensor] = None,

src/transformers/models/video_llava/configuration_video_llava.py

@@ -49,8 +49,10 @@ class VideoLlavaConfig(PretrainedConfig):
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the CLIP backbone.
             Can be either "full" to select all features or "default" to select features without `CLS`.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]]`, *optional*, defaults to -2):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         image_seq_length (`int`, *optional*, defaults to 256):
             Sequence length of one image embedding.
         video_seq_length (`int`, *optional*, defaults to 2056):

src/transformers/models/video_llava/modeling_video_llava.py

@@ -89,8 +89,12 @@ class VideoLlavaCausalLMOutputWithPast(ModelOutput):
 class VideoLlavaMultiModalProjector(nn.Module):
     def __init__(self, config: VideoLlavaConfig):
         super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
         self.linear_1 = nn.Linear(
-            config.vision_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
         )
         self.act = ACT2FN[config.projector_hidden_act]
         self.linear_2 = nn.Linear(
@@ -210,8 +214,10 @@ VIDEO_LLAVA_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
             is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
             model's internal embedding lookup matrix.
-        vision_feature_layer (`int`, *optional*, defaults to -2):
+        vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`):
-            The index of the layer to select the vision feature.
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
         vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
             The feature selection strategy used to select the vision feature from the vision backbone.
             Can be one of `"default"` or `"full"`
@@ -357,7 +363,10 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel, GenerationMi
         return final_embedding, final_attention_mask, final_labels, position_ids, final_input_ids
 
     def get_image_features(
-        self, pixel_values_images: torch.FloatTensor, vision_feature_layer: int, vision_feature_select_strategy: str
+        self,
+        pixel_values_images: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+        vision_feature_select_strategy: str,
     ):
         """
         Obtains image last hidden states from the vision tower and apply multimodal projection.
@@ -365,38 +374,53 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel, GenerationMi
         Args:
             pixel_values_images (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
                The tensors corresponding to the input images.
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
             vision_feature_select_strategy (`str`):
                 The feature selection strategy used to select the vision feature from the vision backbone.
                 Can be one of `"default"` or `"full"`
         Returns:
             image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
         """
+        if vision_feature_select_strategy not in ["default", "full"]:
+            raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}")
 
         image_outputs = self.image_tower(pixel_values_images, output_hidden_states=True)
-        image_outputs = image_outputs.hidden_states[vision_feature_layer].squeeze(1)
 
-        if vision_feature_select_strategy == "default":
+        # If we have one vision feature layer, return the corresponding hidden states,
-            image_outputs = image_outputs[:, 1:]
+        # otherwise, select the hidden states of each feature layer and concatenate them
-        elif vision_feature_select_strategy == "full":
+        if isinstance(vision_feature_layer, int):
-            image_outputs = image_outputs
+            image_outputs = image_outputs.hidden_states[vision_feature_layer]
+            if vision_feature_select_strategy == "default":
+                image_outputs = image_outputs[:, 1:]
         else:
-            raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}")
+            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            # For default; crop CLS from each hidden state in the hidden state pool
+            if vision_feature_select_strategy == "default":
+                hs_pool = [hs[:, 1:] for hs in hs_pool]
+            image_outputs = torch.cat(hs_pool, dim=-1)
 
         image_features = self.multi_modal_projector(image_outputs)
 
         return image_features
 
-    def get_video_features(self, pixel_values_videos: torch.FloatTensor, vision_feature_layer: int):
+    def get_video_features(
+        self,
+        pixel_values_videos: torch.FloatTensor,
+        vision_feature_layer: Union[int, List[int]],
+    ):
         """
         Obtains video last hidden states from the vision tower and apply multimodal projection.
 
         Args:
             pixel_values_videos (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
                The tensors corresponding to the input videos.
-            vision_feature_layer (`int`):
+            vision_feature_layer (`Union[int, List[int]]`):
-                The index of the layer to select the vision feature.
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
         Returns:
             video_features (`torch.Tensor`): Video feature tensor of shape `(num_videos * num_frames, image_length, embed_dim)`).
             frames (`int`): Number of frames the videos have.
@@ -405,7 +429,15 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel, GenerationMi
 
         pixel_values = pixel_values_videos.reshape(batch_size_vid * num_frames, channels, height, width)
         video_outputs = self.video_tower(pixel_values, output_hidden_states=True)
-        video_features = video_outputs.hidden_states[vision_feature_layer].squeeze(1)
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            video_features = video_outputs.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [video_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            video_features = torch.cat(hs_pool, dim=-1)
+
         video_features = self.multi_modal_projector(video_features)
 
         return video_features, num_frames
@@ -422,7 +454,7 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel, GenerationMi
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layer: Optional[int] = None,
+        vision_feature_layer: Optional[Union[int, List[int]]] = None,
         vision_feature_select_strategy: Optional[str] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,

src/transformers/models/vipllava/configuration_vipllava.py

@@ -45,8 +45,8 @@ class VipLlavaConfig(PretrainedConfig):
             The activation function used by the multimodal projector.
         projector_layernorm_eps (`float`, *optional*, defaults to 1e-05):
             The layer norm epsilon of the projector layernorm
-        vision_feature_layers (`List[int]`, *optional*, defaults to `[-2, -5, -8, -11, 6]`):
+        vision_feature_layers (`Union[int, List[int]]`, *optional*, defaults to `[-2, -5, -8, -11, 6]`):
-            The list of layers to select the vision features from.
+            The vision feature layer, or list of layers to select the vision features from.
         image_seq_length (`int`, *optional*, defaults to 576):
             Sequence length of one image embedding.
 

src/transformers/models/vipllava/modeling_vipllava.py

@@ -85,12 +85,13 @@ class VipLlavaCausalLMOutputWithPast(ModelOutput):
 class VipLlavaMultiModalProjector(nn.Module):
     def __init__(self, config: VipLlavaConfig):
         super().__init__()
+        num_feature_layers = 1 if isinstance(config.vision_feature_layers, int) else len(config.vision_feature_layers)
         self.projector_layernorm = nn.LayerNorm(
-            len(config.vision_feature_layers) * config.vision_config.hidden_size, eps=config.projector_layernorm_eps
+            num_feature_layers * config.vision_config.hidden_size, eps=config.projector_layernorm_eps
         )
 
         self.linear_1 = nn.Linear(
-            len(config.vision_feature_layers) * config.vision_config.hidden_size,
+            num_feature_layers * config.vision_config.hidden_size,
             config.text_config.hidden_size,
             bias=True,
         )
@@ -270,24 +271,29 @@ class VipLlavaForConditionalGeneration(VipLlavaPreTrainedModel, GenerationMixin)
         return self.language_model.get_decoder()
 
     # Ignore copy
-    def get_image_features(self, pixel_values: torch.FloatTensor, vision_feature_layers: List[int]):
+    def get_image_features(self, pixel_values: torch.FloatTensor, vision_feature_layers: Union[int, List[int]]):
         """
         Obtains image last hidden states from the vision tower and apply multimodal projection.
 
         Args:
             pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
                The tensors corresponding to the input images.
-            vision_feature_layers (`List[int]`):
+            vision_feature_layers (`Union[int, List[int]]`):
-                The list og indexes of the layers to select the vision feature.
+                The vision feature layer, or the list of indexes of the layers to select
+                the vision feature.
         Returns:
             image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
         """
         image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
 
-        # For VIP-llava, the image features are computed this way
+        # If multiple feature layers are provided (which is usually the case)
-        # We select the features from index 1: for the layers -2, -5, -8, -11 and 6
+        # then the image features are concatenated after the CLS is removed.
-        image_features = [image_outputs.hidden_states[index][:, 1:] for index in vision_feature_layers]
+        if isinstance(vision_feature_layers, int):
-        image_features = torch.cat(image_features, dim=-1)
+            image_features = image_outputs.hidden_states[vision_feature_layers][:, 1:]
+        else:
+            # Usually, we select the features from index 1: the layers -2, -5, -8, -11 and 6
+            image_features = [image_outputs.hidden_states[index][:, 1:] for index in vision_feature_layers]
+            image_features = torch.cat(image_features, dim=-1)
         image_features = self.multi_modal_projector(image_features)
         return image_features
 
@@ -386,7 +392,7 @@ class VipLlavaForConditionalGeneration(VipLlavaPreTrainedModel, GenerationMixin)
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[List[torch.FloatTensor]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
-        vision_feature_layers: Optional[List[int]] = None,
+        vision_feature_layers: Optional[Union[int, List[int]]] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
         output_attentions: Optional[bool] = None,

tests/models/llava/test_modeling_llava.py

@@ -17,6 +17,7 @@
 import unittest
 
 import requests
+from parameterized import parameterized
 
 from transformers import (
     AutoProcessor,
@@ -272,6 +273,32 @@ class LlavaForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterM
             pixel_values = torch.cat([pixel_values, pixel_values], dim=0)
             _ = model(input_ids=input_ids, pixel_values=pixel_values)
 
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layer):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layer = vision_feature_layer
+
+        num_feature_layers = 1 if isinstance(vision_feature_layer, int) else len(vision_feature_layer)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            assert model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
     @unittest.skip(
         reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
     )

tests/models/llava_next/test_modeling_llava_next.py

@@ -18,6 +18,7 @@ import unittest
 
 import requests
 from huggingface_hub import hf_hub_download
+from parameterized import parameterized
 
 from transformers import (
     AutoProcessor,
@@ -321,6 +322,32 @@ class LlavaNextForConditionalGenerationModelTest(ModelTesterMixin, GenerationTes
             image_sizes = torch.cat([image_sizes, image_sizes], dim=0)
             _ = model(input_ids=input_ids, pixel_values=pixel_values, image_sizes=image_sizes)
 
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layer):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layer = vision_feature_layer
+
+        num_feature_layers = 1 if isinstance(vision_feature_layer, int) else len(vision_feature_layer)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            assert model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
     @unittest.skip(
         reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
     )
@@ -558,3 +585,25 @@ class LlavaNextForConditionalGenerationIntegrationTest(unittest.TestCase):
             self.processor.decode(output[0], skip_special_tokens=True),
             EXPECTED_DECODED_TEXT,
         )
+
+    @unittest.skip(reason="Granite multimodal [vision] models are not yet released")
+    @slow
+    def test_granite_vision(self):
+        """
+        Check the expected output of a granite vision model, which leverages
+        multiple vision feature layers and a visual encoder with no CLS (siglip).
+        """
+        # TODO @alex-jw-brooks - update the path and enable this test once the 2b model is released
+        granite_model_path = "llava-granite-2b"
+        model = LlavaNextForConditionalGeneration.from_pretrained(granite_model_path)
+        self.processor = AutoProcessor.from_pretrained(granite_model_path)
+        prompt = "<|user|>\n<image>\nWhat is shown in this image?\n<|assistant|>\n"
+        inputs = self.processor(prompt, self.image, return_tensors="pt").to(model.device)
+
+        # verify generation
+        output = model.generate(**inputs, max_new_tokens=30)
+        EXPECTED_DECODED_TEXT = "<|user|>\n\nWhat is shown in this image?\n<|assistant|>\nThe image depicts a diagram."
+        self.assertEqual(
+            self.processor.decode(output[0], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )

tests/models/llava_next_video/test_modeling_llava_next_video.py

@@ -18,6 +18,7 @@ import unittest
 
 import numpy as np
 from huggingface_hub import hf_hub_download
+from parameterized import parameterized
 
 from transformers import (
     AutoProcessor,
@@ -338,6 +339,32 @@ class LlavaNextVideoForConditionalGenerationModelTest(ModelTesterMixin, Generati
             image_sizes = torch.cat([image_sizes, image_sizes], dim=0)
             _ = model(input_ids=input_ids, pixel_values=pixel_values, image_sizes=image_sizes)
 
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layer):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layer = vision_feature_layer
+
+        num_feature_layers = 1 if isinstance(vision_feature_layer, int) else len(vision_feature_layer)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            assert model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
     @unittest.skip(
         reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
     )

tests/models/llava_onevision/test_modeling_llava_onevision.py

@@ -19,6 +19,7 @@ import unittest
 import numpy as np
 import requests
 from huggingface_hub import hf_hub_download
+from parameterized import parameterized
 
 from transformers import (
     AutoProcessor,
@@ -292,6 +293,32 @@ class LlavaOnevisionForConditionalGenerationModelTest(ModelTesterMixin, Generati
                 out_embeds = model(inputs_embeds=inputs_embeds, **inputs)[0]
             self.assertTrue(torch.allclose(out_embeds, out_ids))
 
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layer):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layer = vision_feature_layer
+
+        num_feature_layers = 1 if isinstance(vision_feature_layer, int) else len(vision_feature_layer)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            assert model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
     @unittest.skip(
         reason="This architecure seem to not compute gradients properly when using GC, SiglipVisionModel does not support standalone training"
     )

tests/models/video_llava/test_modeling_video_llava.py

@@ -19,6 +19,7 @@ import unittest
 import numpy as np
 import requests
 from huggingface_hub import hf_hub_download
+from parameterized import parameterized
 
 from transformers import (
     VideoLlavaConfig,
@@ -419,6 +420,32 @@ class VideoLlavaForConditionalGenerationModelTest(ModelTesterMixin, GenerationTe
             pixel_values = torch.cat([pixel_values, pixel_values], dim=0)
             _ = model(input_ids=input_ids, pixel_values_images=pixel_values)
 
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layer):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layer = vision_feature_layer
+
+        num_feature_layers = 1 if isinstance(vision_feature_layer, int) else len(vision_feature_layer)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            assert model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
 
 @require_torch
 class VideoLlavaForConditionalGenerationIntegrationTest(unittest.TestCase):

tests/models/vipllava/test_modeling_vipllava.py

@@ -17,6 +17,7 @@
 import unittest
 
 import requests
+from parameterized import parameterized
 
 from transformers import (
     AutoProcessor,
@@ -257,6 +258,37 @@ class VipLlavaForConditionalGenerationModelTest(ModelTesterMixin, GenerationTest
             pixel_values = torch.cat([pixel_values, pixel_values], dim=0)
             _ = model(input_ids=input_ids, pixel_values=pixel_values)
 
+    @parameterized.expand(
+        [
+            (-1,),
+            ([-1],),
+            ([-1, -2],),
+        ],
+    )
+    def test_vision_feature_layers(self, vision_feature_layers):
+        """
+        Test that we can use either one vision feature layer, or a list of
+        vision feature layers.
+        """
+        # NOTE: vipllava uses vision_feature_layers instead of vision_feature_layer as the
+        # config key. The reason is that other llava classes supported one vision feature layer
+        # and added support for a list of layers with granite vision support, while vipllava
+        # originally supported multiple feature layers, and added support for a single layer for
+        # for compatibility reasons.
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.vision_feature_layers = vision_feature_layers
+
+        num_feature_layers = 1 if isinstance(vision_feature_layers, int) else len(vision_feature_layers)
+        hidden_size = config.vision_config.hidden_size
+        expected_features = hidden_size * num_feature_layers
+
+        for model_class in self.all_model_classes:
+            model = model_class(config).to(torch_device)
+            # We should have the right number of input features,
+            # and should be able to run a forward pass without exploding
+            assert model.multi_modal_projector.linear_1.in_features == expected_features
+            model(**input_dict)
+
     @unittest.skip(
         reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
     )