[`Llava`] Add Llava to transformers (#27662)

* add model like

* logits match

* minor fixes

* fixes

* up

* up

* add todo

* llava processor

* keep the processor simple

* add conversion script

* fixup

* fix copies

* up

* add to index

* fix config + logits

* fix

* refactor

* more refactor

* more refactor

* fix copies

* add authors

* v1 tests

* add `LlavaProcessor` in init

* remove unneeded import

* up

* up

* docs

* up

* fix CI

* fix CI

* add attention  mask in test

* make fixup

* remove the vision model

* that' s the dirty way to do it

* nits

* nits

* updates

* add more tests

* add input tests

* fixup

* more styling

* nits

* updates amd cleanup

* fixup the generation expected results

* fix the testing script

* some cleanup and simplification which does not work yet but almost there!

* make correct dispatch operations

* vectorize works for batch of images and text

* last todos

* nits

* update test and modeling code

* remove useless function for now

* fix few issues

* fix generation

* some nits

* add bakllava

* nits

* remove duplicated code

* finis merge

* cleanup

* missed this line

* fill the todos

* add left padding offset

* add left and rignt padding logic

* bool to properly index

* make sure

* more cleanups

* batch is fixed 😉



* add correct device for tensor creation

* fix some dtype missmatch

* ruff

* update conversion script

* Update src/transformers/__init__.py

* fa 2 support + fix conversion script

* more

* correct reshaping

* fix test dict

* fix copies by ignoring

* fix nit

* skip clip vision model

* fixup

* fixup

* LlavaForVisionText2Text -> LlavaForCausalLM

* update

* fix

* raise correct errors

* fix

* docs

* nuke for now

* nits here and there

* fixup

* fix remaining tests

* update LlavaForConditionalGeneration instead of CausalLM

* fixups

* pipeline support

* slow and piepline tests

* supports batch

* nits

* cleanup

* fix first integration tests

* add pad token where needed

* correct etsts

* fixups

* update pipeline testr

* fix quality

* nits

* revert unneeded change

* nit

* use BatchFeature

* from ...feature_extraction_utils import BatchFeature

* nits

* nits

* properly update

* more f*** nits

* fix copies

* comment

* keep slow test slow

* Update src/transformers/models/llava/processing_llava.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* add piepline example

* add pixel values in docstrign

* update pr doctest

* fix

* fix slow tests

* remove hack

* fixup

* small note

* forward contrib credits from PR25789

* forward contrib credits from original implementation and work

* add arthur

* Update src/transformers/models/llava/processing_llava.py
Co-authored-by: Lysandre Debut <hi@lysand.re>

* update docstring

* nit

* move to not doctested because of timeout issues

* fixup

* add description

* more

* fix-copies

* fix docs

* add beam search

* add more comments

* add typehints on processor

* add speedup plot

* update slow tests and docs

* push test

* push batched test

* fix batched generation with different number of images

* remove benchmark due to a bug

* fix test

* fix copies

* add gcolab demo

---------
Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: shauray8 <shauray8@users.noreply.github.com>
Co-authored-by: haotian-liu <haotian-liu@users.noreply.github.com>
Co-authored-by: Lysandre Debut <hi@lysand.re>

src/transformers/models/llava/modeling_llava.py

+# coding=utf-8
+# Copyright 2023 the HuggingFace Inc. 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.
+""" PyTorch Llava model."""
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ... import PreTrainedModel
+from ...activations import ACT2FN
+from ...modeling_outputs import ModelOutput
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from ..auto import AutoModel, AutoModelForCausalLM
+from .configuration_llava import LlavaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlavaConfig"
+
+LLAVA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "llava-hf/llava-1.5-7b-hf",
+    "llava-hf/llava-1.5-13b-hf",
+    "llava-hf/bakLlava-v1-hf",
+    # See all Llava models at https://huggingface.co/models?filter=llava
+]
+
+
+@dataclass
+# Copied from transformers.models.idefics.modeling_idefics.IdeficsCausalLMOutputWithPast with Idefics->Llava
+class LlavaCausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for Llava causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
+            Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
+            sequence_length, hidden_size)`.
+
+            image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[List[torch.FloatTensor]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class LlavaMultiModalProjector(nn.Module):
+    def __init__(self, config: LlavaConfig):
+        super().__init__()
+
+        self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
+        self.act = ACT2FN[config.projector_hidden_act]
+        self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
+
+    def forward(self, image_features):
+        hidden_states = self.linear_1(image_features)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+LLAVA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LlavaConfig`] or [`LlavaVisionConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAVA_START_DOCSTRING,
+)
+class LlavaPreTrainedModel(PreTrainedModel):
+    config_class = LlavaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlavaVisionAttention"]
+    _supports_flash_attn_2 = True
+
+    def _init_weights(self, module):
+        # important: this ported version of Llava isn't meant for training from scratch - only
+        # inference and fine-tuning - so the proper init weights code has been removed - the original codebase
+        # https://github.com/haotian-liu/LLaVA/tree/main/llava should serve for that purpose
+        std = (
+            self.config.initializer_range
+            if hasattr(self.config, "initializer_range")
+            else self.config.text_config.initializer_range
+        )
+
+        if hasattr(module, "class_embedding"):
+            module.class_embedding.data.normal_(mean=0.0, std=std)
+
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+LLAVA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
+            The tensors corresponding to the input images. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details ([]`LlavaProcessor`] uses
+            [`CLIPImageProcessor`] for processing images).
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            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.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    """The LLAVA model which consists of a vision backbone and a language model.""",
+    LLAVA_START_DOCSTRING,
+)
+class LlavaForConditionalGeneration(LlavaPreTrainedModel):
+    def __init__(self, config: LlavaConfig):
+        super().__init__(config)
+        self.vision_tower = AutoModel.from_config(config.vision_config)
+
+        self.multi_modal_projector = LlavaMultiModalProjector(config)
+        self.vocab_size = config.vocab_size
+        use_flash_attention_2 = getattr(config, "_flash_attn_2_enabled", False)
+        self.language_model = AutoModelForCausalLM.from_config(
+            config.text_config, use_flash_attention_2=use_flash_attention_2
+        )
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def set_decoder(self, decoder):
+        self.language_model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.language_model.get_decoder()
+
+    def tie_weights(self):
+        return self.language_model.tie_weights()
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
+        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        # update vocab size
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        self.config.vocab_size = model_embeds.num_embeddings
+        self.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+    def _merge_input_ids_with_image_features(
+        self, image_features, inputs_embeds, input_ids, attention_mask, position_ids
+    ):
+        nb_images, image_hidden_dim, embed_dim = image_features.shape
+        batch_size, sequence_length = input_ids.shape
+        left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
+        # 1. Create a mask to know where image tokens are
+        image_token_mask = input_ids == self.config.image_token_index
+        num_image_tokens = torch.sum(image_token_mask, dim=-1)
+        # Compute the maximum embed dimension
+        max_embed_dim = (num_image_tokens.max() * (image_hidden_dim - 1)) + sequence_length
+        batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)
+
+        # 2. Compute the positions where text should be written
+        # Calculate new positions for text tokens in merged image-text sequence.
+        # `image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
+        # `torch.cumsum` computes how each image token shifts subsequent text token positions.
+        # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
+        new_token_positions = torch.cumsum((image_token_mask * (image_hidden_dim - 1) + 1), -1) - 1
+        nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
+        if left_padding:
+            new_token_positions += nb_image_pad[:, None]  # offset for left padding
+        text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
+
+        # 3. Create the full embedding, already padded to the maximum position
+        final_embedding = torch.zeros(
+            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
+        )
+        final_attention_mask = torch.zeros(
+            batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
+        )
+
+        # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
+        # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
+        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
+        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
+
+        # 5. Fill the embeddings corresponding to the images. Anything that is still zeros needs filling
+        image_to_overwrite = torch.all(final_embedding == 0, dim=-1)
+        image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None]
+
+        if image_to_overwrite.sum() != image_features.shape[:-1].numel():
+            raise ValueError(
+                f"The input provided to the model are wrong. The number of image tokens is {torch.sum(image_token_mask)} while"
+                f" the number of image given to the model is {nb_images}. This prevents correct indexing and breaks batch generation."
+            )
+
+        final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim)
+        final_attention_mask |= image_to_overwrite
+        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
+        return final_embedding, final_attention_mask, position_ids
+
+    @add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        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_select_strategy: Optional[str] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+        >>> model = LlavaForConditionalGeneration.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> processor = AutoProcessor.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "<image>\nUSER: What's the content of the image?\nASSISTANT:"
+        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(text=text, images=image, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(**inputs, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "There seems to be a stop sign"
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+
+        if inputs_embeds is None:
+            # 1. Extra the input embeddings
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+            # 2. Merge text and images
+            if pixel_values is not None and input_ids.shape[1] != 1:
+                image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
+                # 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}"
+                    )
+
+                image_features = self.multi_modal_projector(selected_image_feature)
+                inputs_embeds, attention_mask, position_ids = self._merge_input_ids_with_image_features(
+                    image_features, inputs_embeds, input_ids, attention_mask, position_ids
+                )
+                if labels is None:
+                    labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long)
+            else:
+                # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
+                # generation with cache
+                if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
+                    # Retrieve the first layer to inspect the logits and mask out the hidden states
+                    # that are set to 0
+                    first_layer_past_key_value = past_key_values[0][0][:, 0, :, 0]
+                    batch_index, non_attended_tokens = torch.where(first_layer_past_key_value == 0)
+                    # Get the target length
+                    target_seqlen = first_layer_past_key_value.shape[-1] + 1
+
+                    extended_attention_mask = torch.ones(
+                        (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),
+                        dtype=attention_mask.dtype,
+                        device=attention_mask.device,
+                    )
+
+                    # Zero-out the places where we don't need to attend
+                    extended_attention_mask[batch_index, non_attended_tokens] = 0
+
+                    attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)
+                    position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        logits = outputs[0]
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                shift_attention_mask = attention_mask[..., 1:]
+                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return LlavaCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, inputs_embeds=None, pixel_values=None, **kwargs
+    ):
+        # Call `prepare_inputs_for_generation` from the LM
+        model_input = self.language_model.prepare_inputs_for_generation(
+            input_ids, past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        model_input.update({"pixel_values": pixel_values})
+        return model_input
+
+    def _reorder_cache(self, *args, **kwargs):
+        return self.language_model._reorder_cache(*args, **kwargs)

src/transformers/models/llava/processing_llava.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# 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.
+"""
+Processor class for Llava.
+"""
+
+
+import warnings
+from typing import Callable, List, Optional, Union
+
+from ...feature_extraction_utils import BatchFeature
+from ...image_utils import ImageInput
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from ...utils import TensorType
+
+
+class LlavaProcessor(ProcessorMixin):
+    r"""
+    Constructs a Llava processor which wraps a Llava image processor and a Llava tokenizer into a single processor.
+
+    [`LlavaProcessor`] offers all the functionalities of [`CLIPImageProcessor`] and [`LlamaTokenizerFast`]. See the
+    [`~LlavaProcessor.__call__`] and [`~LlavaProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`CLIPImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`LlamaTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "CLIPImageProcessor"
+    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.__init__
+    def __init__(self, image_processor=None, tokenizer=None, **kwargs):
+        feature_extractor = None
+        if "feature_extractor" in kwargs:
+            warnings.warn(
+                "The `feature_extractor` argument is deprecated and will be removed in v5, use `image_processor`"
+                " instead.",
+                FutureWarning,
+            )
+            feature_extractor = kwargs.pop("feature_extractor")
+
+        image_processor = image_processor if image_processor is not None else feature_extractor
+        if image_processor is None:
+            raise ValueError("You need to specify an `image_processor`.")
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        images: ImageInput = None,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        transform: Callable = None,
+        max_length=None,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            truncation (`bool`, *optional*):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            transform (`Callable`, *optional*):
+                A custom transform function that accepts a single image can be passed for training. For example,
+                `torchvision.Compose` can be used to compose multiple functions. If `None` a preset inference-specific
+                set of transforms will be applied to the images
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if images is not None:
+            pixel_values = self.image_processor(images, transform=transform, return_tensors=return_tensors)[
+                "pixel_values"
+            ]
+        else:
+            pixel_values = None
+        text_inputs = self.tokenizer(
+            text, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length
+        )
+
+        return BatchFeature(data={**text_inputs, "pixel_values": pixel_values})
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

src/transformers/utils/dummy_pt_objects.py

@@ -4641,6 +4641,30 @@ class LlamaPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+LLAVA_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class LlavaForConditionalGeneration(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class LlavaPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class LlavaProcessor(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/llava/test_modeling_llava.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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.
+""" Testing suite for the PyTorch Llava model. """
+
+import gc
+import unittest
+
+import requests
+
+from transformers import (
+    AutoProcessor,
+    LlavaConfig,
+    LlavaForConditionalGeneration,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import require_bitsandbytes, require_torch, slow, torch_device
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+else:
+    is_torch_greater_or_equal_than_2_0 = False
+
+if is_vision_available():
+    from PIL import Image
+
+
+class LlavaVisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        ignore_index=-100,
+        image_token_index=0,
+        projector_hidden_act="gelu",
+        seq_length=7,
+        vision_feature_select_strategy="default",
+        vision_feature_layer=-1,
+        text_config={
+            "model_type": "llama",
+            "seq_length": 7,
+            "is_training": True,
+            "use_input_mask": True,
+            "use_token_type_ids": False,
+            "use_labels": True,
+            "vocab_size": 99,
+            "hidden_size": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 37,
+            "hidden_act": "gelu",
+            "hidden_dropout_prob": 0.1,
+            "attention_probs_dropout_prob": 0.1,
+            "max_position_embeddings": 512,
+            "type_vocab_size": 16,
+            "type_sequence_label_size": 2,
+            "initializer_range": 0.02,
+            "num_labels": 3,
+            "num_choices": 4,
+            "pad_token_id": 0,
+        },
+        is_training=True,
+        vision_config={
+            "batch_size": 12,
+            "image_size": 30,
+            "patch_size": 2,
+            "num_channels": 3,
+            "is_training": True,
+            "hidden_size": 32,
+            "projection_dim": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 37,
+            "dropout": 0.1,
+            "attention_dropout": 0.1,
+            "initializer_range": 0.02,
+        },
+    ):
+        self.parent = parent
+        self.ignore_index = ignore_index
+        self.image_token_index = image_token_index
+        self.projector_hidden_act = projector_hidden_act
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.vision_feature_layer = vision_feature_layer
+        self.text_config = text_config
+        self.vision_config = vision_config
+        self.seq_length = seq_length
+
+        self.num_hidden_layers = text_config["num_hidden_layers"]
+        self.vocab_size = text_config["vocab_size"]
+        self.hidden_size = text_config["hidden_size"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+        self.is_training = is_training
+
+        self.batch_size = 3
+        self.num_channels = 3
+        self.image_size = 336
+        self.encoder_seq_length = 231
+
+    def get_config(self):
+        return LlavaConfig(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            ignore_index=self.ignore_index,
+            image_token_index=self.image_token_index,
+            projector_hidden_act=self.projector_hidden_act,
+            vision_feature_select_strategy=self.vision_feature_select_strategy,
+            vision_feature_layer=self.vision_feature_layer,
+        )
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [
+                self.batch_size,
+                self.vision_config["num_channels"],
+                self.vision_config["image_size"],
+                self.vision_config["image_size"],
+            ]
+        )
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 1) + 1
+        attention_mask = input_ids.ne(1).to(torch_device)
+        # we are giving 3 images let's make sure we pass in 3 image tokens
+        input_ids[:, 1] = config.image_token_index
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class LlavaForConditionalGenerationModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Model tester for `LlavaForConditionalGeneration`.
+    """
+
+    all_model_classes = (LlavaForConditionalGeneration,) if is_torch_available() else ()
+    pipeline_model_mapping = {"image-to-text": LlavaForConditionalGeneration} if is_torch_available() else {}
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = True
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = LlavaVisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=LlavaConfig, has_text_modality=False)
+
+    @unittest.skip(
+        reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(
+        reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        pass
+
+    @unittest.skip(
+        reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+
+@require_torch
+class LlavaForConditionalGenerationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("llava-hf/bakLlava-v1-hf")
+
+    def tearDown(self):
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test(self):
+        # Let' s make sure we test the preprocessing to replace what is used
+        model = LlavaForConditionalGeneration.from_pretrained("llava-hf/bakLlava-v1-hf", load_in_4bit=True)
+
+        prompt = "<image>\nUSER: What are the things I should be cautious about when I visit this place?\nASSISTANT:"
+        image_file = "https://llava-vl.github.io/static/images/view.jpg"
+        raw_image = Image.open(requests.get(image_file, stream=True).raw)
+        inputs = self.processor(prompt, raw_image, return_tensors="pt")
+
+        EXPECTED_INPUT_IDS = torch.tensor([[1, 32000, 28705, 13, 11123, 28747, 1824, 460, 272, 1722,315, 1023, 347, 13831, 925, 684, 739, 315, 3251, 456,1633, 28804, 13, 4816, 8048, 12738, 28747]])  # fmt: skip
+        self.assertTrue(torch.equal(inputs["input_ids"], EXPECTED_INPUT_IDS))
+
+        output = model.generate(**inputs, max_new_tokens=20)
+        EXPECTED_DECODED_TEXT = "\nUSER: What are the things I should be cautious about when I visit this place?\nASSISTANT: When visiting this place, there are a few things one should be cautious about. Firstly,"  # fmt: skip
+
+        self.assertEqual(
+            self.processor.decode(output[0], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test_llama(self):
+        # Let' s make sure we test the preprocessing to replace what is used
+        model_id = "llava-hf/llava-1.5-7b-hf"
+
+        model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf", load_in_4bit=True)
+        processor = AutoProcessor.from_pretrained(model_id)
+
+        prompt = "USER: <image>\nWhat are the things I should be cautious about when I visit this place?\nASSISTANT:"
+        image_file = "https://llava-vl.github.io/static/images/view.jpg"
+        raw_image = Image.open(requests.get(image_file, stream=True).raw)
+        inputs = processor(prompt, raw_image, return_tensors="pt").to(torch_device, torch.float16)
+
+        output = model.generate(**inputs, max_new_tokens=900, do_sample=False)
+        EXPECTED_DECODED_TEXT = "USER:  \nWhat are the things I should be cautious about when I visit this place?\nASSISTANT: When visiting this place, which is a pier or dock extending over a body of water, there are a few things to be cautious about. First, be aware of the weather conditions, as sudden changes in weather can make the pier unsafe to walk on. Second, be mindful of the water depth and any potential hazards, such as submerged rocks or debris, that could cause accidents or injuries. Additionally, be cautious of the presence of wildlife, such as birds or fish, and avoid disturbing their natural habitats. Lastly, be aware of any local regulations or guidelines for the use of the pier, as some areas may be restricted or prohibited for certain activities."  # fmt: skip
+
+        self.assertEqual(
+            processor.decode(output[0], skip_special_tokens=True),
+            EXPECTED_DECODED_TEXT,
+        )
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test_llama_batched(self):
+        # Let' s make sure we test the preprocessing to replace what is used
+        model_id = "llava-hf/llava-1.5-7b-hf"
+
+        model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf", load_in_4bit=True)
+        processor = AutoProcessor.from_pretrained(model_id, pad_token="<pad>")
+
+        prompts = [
+            "USER: <image>\nWhat are the things I should be cautious about when I visit this place? What should I bring with me?\nASSISTANT:",
+            "USER: <image>\nWhat is this?\nASSISTANT: Two cats lying on a bed!\nUSER: <image>\nAnd this?\nASSISTANT:",
+        ]
+        image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
+        image2 = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = processor(prompts, images=[image1, image2, image1], return_tensors="pt", padding=True)
+
+        output = model.generate(**inputs, max_new_tokens=20)
+
+        EXPECTED_DECODED_TEXT = ['USER:  \nWhat are the things I should be cautious about when I visit this place? What should I bring with me?\nASSISTANT: the water is calm and clear\n\nThe image shows a wooden pier on a lake, with a', 'USER:  \nWhat is this?\nASSISTANT: Two cats lying on a bed!\nUSER:  \nAnd this?\nASSISTANT: A cat sleeping on a bed.']  # fmt: skip
+
+        self.assertEqual(processor.batch_decode(output, skip_special_tokens=True), EXPECTED_DECODED_TEXT)
+
+    @slow
+    @require_bitsandbytes
+    def test_small_model_integration_test_batch(self):
+        # Let' s make sure we test the preprocessing to replace what is used
+        model = LlavaForConditionalGeneration.from_pretrained("llava-hf/bakLlava-v1-hf", load_in_4bit=True)
+        # The first batch is longer in terms of text, but only has 1 image. The second batch will be padded in text, but the first will be padded because images take more space!.
+        prompts = [
+            "USER: <image>\nWhat are the things I should be cautious about when I visit this place? What should I bring with me?\nASSISTANT:",
+            "USER: <image>\nWhat is this?\nASSISTANT: Two cats lying on a bed!\nUSER: <image>\nAnd this?\nASSISTANT:",
+        ]
+        image1 = Image.open(requests.get("https://llava-vl.github.io/static/images/view.jpg", stream=True).raw)
+        image2 = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = self.processor(prompts, images=[image1, image2, image1], return_tensors="pt", padding=True)
+
+        output = model.generate(**inputs, max_new_tokens=20)
+
+        EXPECTED_DECODED_TEXT = ['\nUSER: What are the things I should be cautious about when I visit this place? What should I bring with me\nASSISTANT: When visiting this place, bring a camera, Park rules may apply, Per person, Sunrise over', '\nUSER: What is this?\nASSISTANT: Two cats lying on a bed!\nUSER: And this? a dock on a lake with two cats on it (Photo credit: Jocelyn R.']  # fmt: skip
+        self.assertEqual(self.processor.batch_decode(output, skip_special_tokens=True), EXPECTED_DECODED_TEXT)

tests/pipelines/test_pipelines_image_to_text.py

@@ -252,3 +252,24 @@ class ImageToTextPipelineTests(unittest.TestCase):
                 [{"generated_text": "a cat laying on a blanket next to a cat laying on a bed "}],
             ],
         )
+
+    @slow
+    @require_torch
+    def test_conditional_generation_llava(self):
+        pipe = pipeline("image-to-text", model="llava-hf/bakLlava-v1-hf")
+        url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/ai2d-demo.jpg"
+        image = Image.open(requests.get(url, stream=True).raw)
+
+        prompt = (
+            "<image>\nUSER: What does the label 15 represent? (1) lava (2) core (3) tunnel (4) ash cloud?\nASSISTANT:"
+        )
+
+        outputs = pipe(image, prompt=prompt, generate_kwargs={"max_new_tokens": 200})
+        self.assertEqual(
+            outputs,
+            [
+                {
+                    "generated_text": "<image> \nUSER: What does the label 15 represent? (1) lava (2) core (3) tunnel (4) ash cloud?\nASSISTANT: Lava"
+                }
+            ],
+        )

utils/check_copies.py

@@ -673,6 +673,7 @@ MODELS_NOT_IN_README = [
     "TimmBackbone",
     "Vision Encoder decoder",
     "VisionTextDualEncoder",
+    "CLIPVisionModel",
 ]
 
 # Template for new entries to add in the main README when we have missing models.

utils/check_table.py

@@ -171,6 +171,7 @@ MODEL_NAMES_WITH_SAME_CONFIG = {
     "XLS-R": "Wav2Vec2",
     "XLSR-Wav2Vec2": "Wav2Vec2",
 }
+MODEL_NAMES_TO_IGNORE = ["CLIPVisionModel"]
 
 
 def get_model_table_from_auto_modules() -> str:
@@ -243,6 +244,8 @@ def get_model_table_from_auto_modules() -> str:
     check = {True: "✅", False: "❌"}
 
     for name in model_names:
+        if name in MODEL_NAMES_TO_IGNORE:
+            continue
         if name in MODEL_NAMES_WITH_SAME_CONFIG.keys():
             prefix = model_name_to_prefix[MODEL_NAMES_WITH_SAME_CONFIG[name]]
         else:

utils/not_doctested.txt

@@ -146,6 +146,7 @@ docs/source/en/model_doc/levit.md
 docs/source/en/model_doc/lilt.md
 docs/source/en/model_doc/llama.md
 docs/source/en/model_doc/llama2.md
+docs/source/en/model_doc/llava.md
 docs/source/en/model_doc/longformer.md
 docs/source/en/model_doc/longt5.md
 docs/source/en/model_doc/luke.md
@@ -294,7 +295,7 @@ docs/source/en/serialization.md
 docs/source/en/tasks/asr.md
 docs/source/en/tasks/audio_classification.md
 docs/source/en/tasks/document_question_answering.md
-docs/source/en/tasks/idefics.md  # causes other tests to fail
+docs/source/en/tasks/idefics.md
 docs/source/en/tasks/image_captioning.md
 docs/source/en/tasks/image_classification.md
 docs/source/en/tasks/language_modeling.md
@@ -432,7 +433,7 @@ src/transformers/models/blip/modeling_blip_text.py
 src/transformers/models/blip/modeling_tf_blip_text.py
 src/transformers/models/blip_2/configuration_blip_2.py
 src/transformers/models/blip_2/convert_blip_2_original_to_pytorch.py
-src/transformers/models/blip_2/modeling_blip_2.py  # causes other tests to fail
+src/transformers/models/blip_2/modeling_blip_2.py
 src/transformers/models/bloom/convert_bloom_original_checkpoint_to_pytorch.py
 src/transformers/models/bloom/modeling_bloom.py
 src/transformers/models/bloom/modeling_flax_bloom.py
@@ -634,6 +635,8 @@ src/transformers/models/lilt/configuration_lilt.py
 src/transformers/models/llama/configuration_llama.py
 src/transformers/models/llama/convert_llama_weights_to_hf.py
 src/transformers/models/llama/modeling_llama.py
+src/transformers/models/llava/configuration_llava.py
+src/transformers/models/llava/modeling_llava.py
 src/transformers/models/longformer/configuration_longformer.py
 src/transformers/models/longformer/convert_longformer_original_pytorch_lightning_to_pytorch.py
 src/transformers/models/longt5/configuration_longt5.py