Add InstructBLIP (#23460)

* Squash 88 commits

* Use markdown

* Remove mdx files due to bad rebase

* Fix modeling files due to bad rebase

* Fix style

* Update comment

* fix

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/models/blip_2/processing_blip_2.py

@@ -18,6 +18,7 @@ Processor class for BLIP-2.
 
 from typing import List, Optional, Union
 
+from ...image_utils import ImageInput
 from ...processing_utils import ProcessorMixin
 from ...tokenization_utils_base import BatchEncoding, PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
 from ...utils import TensorType
@@ -49,7 +50,7 @@ class Blip2Processor(ProcessorMixin):
     # Copied from transformers.models.blip.processing_blip.BlipProcessor.__call__
     def __call__(
         self,
-        images=None,
+        images: ImageInput = None,
         text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
         add_special_tokens: bool = True,
         padding: Union[bool, str, PaddingStrategy] = False,

src/transformers/models/instructblip/__init__.py

+# Copyright 2023 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.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_instructblip": [
+        "INSTRUCTBLIP_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "InstructBlipConfig",
+        "InstructBlipQFormerConfig",
+        "InstructBlipVisionConfig",
+    ],
+    "processing_instructblip": ["InstructBlipProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_instructblip"] = [
+        "INSTRUCTBLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "InstructBlipQFormerModel",
+        "InstructBlipPreTrainedModel",
+        "InstructBlipForConditionalGeneration",
+        "InstructBlipVisionModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_instructblip import (
+        INSTRUCTBLIP_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        InstructBlipConfig,
+        InstructBlipQFormerConfig,
+        InstructBlipVisionConfig,
+    )
+    from .processing_instructblip import InstructBlipProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_instructblip import (
+            INSTRUCTBLIP_PRETRAINED_MODEL_ARCHIVE_LIST,
+            InstructBlipForConditionalGeneration,
+            InstructBlipPreTrainedModel,
+            InstructBlipQFormerModel,
+            InstructBlipVisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/instructblip/configuration_instructblip.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.
+""" InstructBLIP model configuration"""
+
+import copy
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+from ...utils import logging
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+INSTRUCTBLIP_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "Salesforce/instruct-blip-flan-t5": "https://huggingface.co/Salesforce/instruct-blip-flan-t5/resolve/main/config.json",
+}
+
+
+class InstructBlipVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`InstructBlipVisionModel`]. It is used to
+    instantiate a InstructBLIP vision encoder according to the specified arguments, defining the model architecture.
+    Instantiating a configuration defaults will yield a similar configuration to that of the InstructBLIP
+    [Salesforce/instruct-blip-flan-t5](https://huggingface.co/Salesforce/instruct-blip-flan-t5) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 1408):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 6144):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 39):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported. to 1e-5): The epsilon used by the layer
+            normalization layers.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-6):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 1e-10):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries and values in the self-attention layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import InstructBlipVisionConfig, InstructBlipVisionModel
+
+    >>> # Initializing a InstructBlipVisionConfig with Salesforce/instruct-blip-flan-t5 style configuration
+    >>> configuration = InstructBlipVisionConfig()
+
+    >>> # Initializing a InstructBlipVisionModel (with random weights) from the Salesforce/instruct-blip-flan-t5 style configuration
+    >>> model = InstructBlipVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "instructblip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=1408,
+        intermediate_size=6144,
+        num_hidden_layers=39,
+        num_attention_heads=16,
+        image_size=224,
+        patch_size=14,
+        hidden_act="gelu",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        initializer_range=1e-10,
+        qkv_bias=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.qkv_bias = qkv_bias
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from InstructBlipConfig
+        if config_dict.get("model_type") == "instructblip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class InstructBlipQFormerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`InstructBlipQFormerModel`]. It is used to
+    instantiate a InstructBLIP Querying Transformer (Q-Former) model according to the specified arguments, defining the
+    model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
+    the InstructBLIP [Salesforce/instruct-blip-flan-t5](https://huggingface.co/Salesforce/instruct-blip-flan-t5)
+    architecture. Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs.
+    Read the documentation from [`PretrainedConfig`] for more information.
+
+    Note that [`InstructBlipQFormerModel`] is very similar to [`BertLMHeadModel`] with interleaved cross-attention.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the Q-Former model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling the model.
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        cross_attention_frequency (`int`, *optional*, defaults to 2):
+            The frequency of adding cross-attention to the Transformer layers.
+        encoder_hidden_size (`int`, *optional*, defaults to 1408):
+            The hidden size of the hidden states for cross-attention.
+
+    Examples:
+
+    ```python
+    >>> from transformers import InstructBlipQFormerConfig, InstructBlipQFormerModel
+
+    >>> # Initializing a InstructBLIP Salesforce/instruct-blip-flan-t5 style configuration
+    >>> configuration = InstructBlipQFormerConfig()
+
+    >>> # Initializing a model (with random weights) from the Salesforce/instruct-blip-flan-t5 style configuration
+    >>> model = InstructBlipQFormerModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "instructblip_qformer"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        cross_attention_frequency=2,
+        encoder_hidden_size=1408,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.cross_attention_frequency = cross_attention_frequency
+        self.encoder_hidden_size = encoder_hidden_size
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the qformer config dict if we are loading from InstructBlipConfig
+        if config_dict.get("model_type") == "instructblip":
+            config_dict = config_dict["qformer_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class InstructBlipConfig(PretrainedConfig):
+    r"""
+    [`InstructBlipConfig`] is the configuration class to store the configuration of a
+    [`InstructBlipForConditionalGeneration`]. It is used to instantiate a InstructBLIP model according to the specified
+    arguments, defining the vision model, Q-Former model and language model configs. Instantiating a configuration with
+    the defaults will yield a similar configuration to that of the InstructBLIP
+    [Salesforce/instruct-blip-flan-t5](https://huggingface.co/Salesforce/instruct-blip-flan-t5) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`InstructBlipVisionConfig`].
+        qformer_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`InstructBlipQFormerConfig`].
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize any [`PretrainedConfig`].
+        num_query_tokens (`int`, *optional*, defaults to 32):
+            The number of query tokens passed through the Transformer.
+
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import (
+    ...     InstructBlipVisionConfig,
+    ...     InstructBlipQFormerConfig,
+    ...     OPTConfig,
+    ...     InstructBlipConfig,
+    ...     InstructBlipForConditionalGeneration,
+    ... )
+
+    >>> # Initializing a InstructBlipConfig with Salesforce/instruct-blip-flan-t5 style configuration
+    >>> configuration = InstructBlipConfig()
+
+    >>> # Initializing a InstructBlipForConditionalGeneration (with random weights) from the Salesforce/instruct-blip-flan-t5 style configuration
+    >>> model = InstructBlipForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a InstructBlipConfig from a InstructBlipVisionConfig, InstructBlipQFormerConfig and any PretrainedConfig
+
+    >>> # Initializing InstructBLIP vision, InstructBLIP Q-Former and language model configurations
+    >>> vision_config = InstructBlipVisionConfig()
+    >>> qformer_config = InstructBlipQFormerConfig()
+    >>> text_config = OPTConfig()
+
+    >>> config = InstructBlipConfig.from_text_vision_configs(vision_config, qformer_config, text_config)
+    ```"""
+
+    model_type = "instructblip"
+    is_composition = True
+
+    def __init__(self, vision_config=None, qformer_config=None, text_config=None, num_query_tokens=32, **kwargs):
+        super().__init__(**kwargs)
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. initializing the InstructBlipVisionConfig with default values.")
+
+        if qformer_config is None:
+            qformer_config = {}
+            logger.info("qformer_config is None. Initializing the InstructBlipQFormerConfig with default values.")
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the text config with default values (`OPTConfig`).")
+
+        self.vision_config = InstructBlipVisionConfig(**vision_config)
+        self.qformer_config = InstructBlipQFormerConfig(**qformer_config)
+        text_model_type = text_config["model_type"] if "model_type" in text_config else "opt"
+        self.text_config = CONFIG_MAPPING[text_model_type](**text_config)
+
+        self.tie_word_embeddings = self.text_config.tie_word_embeddings
+        self.is_encoder_decoder = self.text_config.is_encoder_decoder
+
+        self.num_query_tokens = num_query_tokens
+        self.qformer_config.encoder_hidden_size = self.vision_config.hidden_size
+        self.use_decoder_only_language_model = self.text_config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+        self.initializer_factor = 1.0
+        self.initializer_range = 0.02
+
+    @classmethod
+    def from_vision_qformer_text_configs(
+        cls,
+        vision_config: InstructBlipVisionConfig,
+        qformer_config: InstructBlipQFormerConfig,
+        text_config: PretrainedConfig,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`InstructBlipConfig`] (or a derived class) from a InstructBLIP vision model, Q-Former and
+        language model configurations.
+
+        Returns:
+            [`InstructBlipConfig`]: An instance of a configuration object
+        """
+
+        return cls(
+            vision_config=vision_config.to_dict(),
+            qformer_config=qformer_config.to_dict(),
+            text_config=text_config.to_dict(),
+            **kwargs,
+        )
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["vision_config"] = self.vision_config.to_dict()
+        output["qformer_config"] = self.qformer_config.to_dict()
+        output["text_config"] = self.text_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/instructblip/convert_instructblip_original_to_pytorch.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.
+"""
+Convert InstructBLIP checkpoints from the original repository.
+
+URL: https://github.com/salesforce/LAVIS/tree/main/projects/instructblip
+"""
+
+import argparse
+
+import requests
+import torch
+
+# pip3 install salesforce-lavis
+# I'm actually installing a slightly modified version: pip3 install git+https://github.com/nielsrogge/LAVIS.git@fix_lavis_float32 (there's also the fix_lavis branch)
+# also note: to convert Vicuna checkpoints, we had to include /home/niels/python_projects/checkpoints/FastChat/vicuna-7b in lavis/configs/models/blip2/blip2_instruct_vicuna7b.yaml
+# same for Vicuna-13b
+from lavis.models import load_model_and_preprocess
+from PIL import Image
+
+from transformers import (
+    AutoTokenizer,
+    BlipImageProcessor,
+    InstructBlipConfig,
+    InstructBlipForConditionalGeneration,
+    InstructBlipProcessor,
+    InstructBlipQFormerConfig,
+    InstructBlipVisionConfig,
+    LlamaConfig,
+    LlamaTokenizerFast,
+    T5Config,
+    T5TokenizerFast,
+)
+from transformers.utils.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
+
+
+def load_demo_image():
+    url = "https://raw.githubusercontent.com/salesforce/LAVIS/main/docs/_static/Confusing-Pictures.jpg"
+    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+    return image
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def create_rename_keys(config):
+    rename_keys = []
+    # fmt: off
+
+    # vision encoder
+    rename_keys.append(("visual_encoder.cls_token", "vision_model.embeddings.class_embedding"))
+    rename_keys.append(("visual_encoder.pos_embed", "vision_model.embeddings.position_embedding"))
+    rename_keys.append(("visual_encoder.patch_embed.proj.weight", "vision_model.embeddings.patch_embedding.weight"))
+    rename_keys.append(("visual_encoder.patch_embed.proj.bias", "vision_model.embeddings.patch_embedding.bias"))
+    rename_keys.append(("ln_vision.weight", "vision_model.post_layernorm.weight"))
+    rename_keys.append(("ln_vision.bias", "vision_model.post_layernorm.bias"))
+
+    for i in range(config.vision_config.num_hidden_layers):
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm1.weight", f"vision_model.encoder.layers.{i}.layer_norm1.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm1.bias", f"vision_model.encoder.layers.{i}.layer_norm1.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm2.weight", f"vision_model.encoder.layers.{i}.layer_norm2.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm2.bias", f"vision_model.encoder.layers.{i}.layer_norm2.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.qkv.weight", f"vision_model.encoder.layers.{i}.self_attn.qkv.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.proj.weight", f"vision_model.encoder.layers.{i}.self_attn.projection.weight",))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.proj.bias", f"vision_model.encoder.layers.{i}.self_attn.projection.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc1.weight", f"vision_model.encoder.layers.{i}.mlp.fc1.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc1.bias", f"vision_model.encoder.layers.{i}.mlp.fc1.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc2.weight", f"vision_model.encoder.layers.{i}.mlp.fc2.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc2.bias", f"vision_model.encoder.layers.{i}.mlp.fc2.bias"))
+
+    # QFormer
+    rename_keys.append(("Qformer.bert.embeddings.LayerNorm.weight", "qformer.embeddings.layernorm.weight"))
+    rename_keys.append(("Qformer.bert.embeddings.LayerNorm.bias", "qformer.embeddings.layernorm.bias"))
+
+    # fmt: on
+    return rename_keys
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+def read_in_q_v_bias(state_dict, config):
+    for i in range(config.vision_config.num_hidden_layers):
+        # read in original q and v biases
+        q_bias = state_dict.pop(f"visual_encoder.blocks.{i}.attn.q_bias")
+        v_bias = state_dict.pop(f"visual_encoder.blocks.{i}.attn.v_bias")
+
+        # next, set bias in the state dict
+        qkv_bias = torch.cat((q_bias, torch.zeros_like(v_bias, requires_grad=False), v_bias))
+        state_dict[f"vision_model.encoder.layers.{i}.self_attn.qkv.bias"] = qkv_bias
+
+
+def get_blip2_config(model_name):
+    image_size = 364 if "coco" in model_name else 224
+    vision_config = InstructBlipVisionConfig(image_size=image_size).to_dict()
+
+    # make sure the models have proper bos_token_id and eos_token_id set (important for generation)
+    # seems like flan-T5 models don't have bos_token_id properly set?
+    if "t5-xl" in model_name:
+        text_config = T5Config.from_pretrained("google/flan-t5-xl", dense_act_fn="gelu", bos_token_id=1).to_dict()
+    elif "t5-xxl" in model_name:
+        text_config = T5Config.from_pretrained("google/flan-t5-xxl", dense_act_fn="gelu", bos_token_id=1).to_dict()
+    elif "vicuna-7b" in model_name:
+        text_config = LlamaConfig.from_pretrained("decapoda-research/llama-7b-hf", vocab_size=32001).to_dict()
+    elif "vicuna-13b" in model_name:
+        text_config = LlamaConfig.from_pretrained("decapoda-research/llama-13b-hf", vocab_size=32001).to_dict()
+    else:
+        raise ValueError("Model name not supported")
+
+    # the authors add one special "[DEC]" token to the vocab of Q-Former, hence vocab size = 30522 + 1
+    qformer_config = InstructBlipQFormerConfig(vocab_size=30523).to_dict()
+    config = InstructBlipConfig(vision_config=vision_config, text_config=text_config, qformer_config=qformer_config)
+
+    return config, image_size
+
+
+@torch.no_grad()
+def convert_blip2_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to Transformers design.
+    """
+    qformer_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased", truncation_side="left")
+    qformer_tokenizer.add_special_tokens({"bos_token": "[DEC]"})
+
+    if "t5" in model_name:
+        tokenizer = T5TokenizerFast.from_pretrained("google/flan-t5-xl", truncation_side="left")
+    elif "vicuna" in model_name:
+        # the following was used in the original implementation:
+        # tokenizer = LlamaTokenizer.from_pretrained("huggyllama/llama-7b", use_fast=False, truncation_side="left")
+        # tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+        # tokenizer.add_special_tokens({"bos_token": "</s>"})
+        # tokenizer.add_special_tokens({"eos_token": "</s>"})
+        # tokenizer.add_special_tokens({"unk_token": "</s>"})
+        tokenizer = LlamaTokenizerFast.from_pretrained(
+            "huggyllama/llama-7b", truncation_side="left", bos_token="</s>", unk_token="</s>"
+        )
+        tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+
+    config, image_size = get_blip2_config(model_name)
+    hf_model = InstructBlipForConditionalGeneration(config).eval()
+
+    model_name_to_original = {
+        "instructblip-vicuna-7b": ("blip2_vicuna_instruct", "vicuna7b"),
+        "instructblip-vicuna-13b": ("blip2_vicuna_instruct", "vicuna13b"),
+        "instructblip-flan-t5-xl": ("blip2_t5_instruct", "flant5xl"),
+        "instructblip-flan-t5-xxl": ("blip2_t5_instruct", "flant5xxl"),
+    }
+
+    name, type = model_name_to_original[model_name]
+
+    # load original model
+    print("Loading original model...")
+    hf_model_device = "cuda:1" if torch.cuda.is_available() else "cpu"
+    lavis_device = "cuda:2" if torch.cuda.is_available() else "cpu"
+    original_model, vis_processors, _ = load_model_and_preprocess(
+        name=name, model_type=type, is_eval=True, device=lavis_device
+    )
+    original_model.eval()
+    print("Done!")
+
+    # update state dict keys
+    state_dict = original_model.state_dict()
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+
+    # some keys can be renamed efficiently
+    for key, val in state_dict.copy().items():
+        val = state_dict.pop(key)
+        if key.startswith("Qformer.bert"):
+            key = key.replace("Qformer.bert", "qformer")
+        if "attention.self" in key:
+            key = key.replace("self", "attention")
+        if "llm_proj" in key:
+            key = key.replace("llm_proj", "language_projection")
+        if "t5_proj" in key:
+            key = key.replace("t5_proj", "language_projection")
+        if key.startswith("llm_model"):
+            key = key.replace("llm_model", "language_model")
+        if key.startswith("t5"):
+            key = key.replace("t5", "language")
+        state_dict[key] = val
+
+    # read in qv biases
+    read_in_q_v_bias(state_dict, config)
+
+    # note: weights get loaded in torch.float32 by default
+    hf_model.load_state_dict(state_dict, strict=True)
+
+    image = load_demo_image()
+    prompt = "What is unusual about this image?"
+
+    # create processor
+    image_processor = BlipImageProcessor(
+        size={"height": image_size, "width": image_size}, image_mean=OPENAI_CLIP_MEAN, image_std=OPENAI_CLIP_STD
+    )
+    processor = InstructBlipProcessor(
+        image_processor=image_processor,
+        tokenizer=tokenizer,
+        qformer_tokenizer=qformer_tokenizer,
+    )
+    inputs = processor(images=image, text=prompt, return_tensors="pt").to(hf_model_device)
+
+    # make sure processor creates exact same pixel values
+    original_pixel_values = vis_processors["eval"](image).unsqueeze(0).to(lavis_device)
+    pixel_values = inputs.pixel_values
+    assert torch.allclose(original_pixel_values.to(pixel_values.device), pixel_values)
+
+    original_model.to(lavis_device)
+    hf_model.to(hf_model_device)
+    with torch.no_grad():
+        if "vicuna" in model_name:
+            original_logits = original_model({"image": original_pixel_values, "text_input": [prompt]}).logits
+            logits = hf_model(**inputs).logits
+        else:
+            original_logits = original_model(
+                {"image": original_pixel_values, "text_input": [prompt], "text_output": ["\n"]}
+            ).logits
+            label_input_ids = tokenizer("\n", return_tensors="pt").input_ids.to(hf_model_device)
+            labels = label_input_ids.masked_fill(label_input_ids == tokenizer.pad_token_id, -100)
+            logits = hf_model(**inputs, labels=labels).logits
+
+    print("First values of original logits:", original_logits[0, :3, :3])
+    print("First values of HF logits:", logits[0, :3, :3])
+
+    # assert values
+    assert original_logits.shape == logits.shape
+    atol = 1e-4 if "vicuna" in model_name else 1e-5
+    assert torch.allclose(original_logits.to(logits.device), logits, atol=atol)
+    print("Looks ok!")
+
+    print("Generating with original model...")
+    original_outputs = original_model.generate({"image": original_pixel_values, "prompt": prompt}, num_beams=5)
+
+    # important: we need to cast the weights of the HF model to the appropriate type
+    print("Generating with HF model...")
+    outputs = hf_model.generate(
+        **inputs,
+        do_sample=False,
+        num_beams=5,
+        max_length=256,
+        min_length=1,
+        top_p=0.9,
+        repetition_penalty=1.5,
+        length_penalty=1.0,
+        temperature=1,
+    )
+    if "vicuna" in model_name:
+        # convert output id 0 to 2 (eos_token_id)
+        # TODO add this in the generate method?
+        outputs[outputs == 0] = 2
+    print("Original generation:", original_outputs)
+    output_text = processor.batch_decode(outputs, skip_special_tokens=True)
+    output_text = [text.strip() for text in output_text]
+    print("HF generation:", output_text)
+
+    if pytorch_dump_folder_path is not None:
+        processor.save_pretrained(pytorch_dump_folder_path)
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        processor.push_to_hub(f"Salesforce/{model_name}")
+        hf_model.push_to_hub(f"Salesforce/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    choices = [
+        "instructblip-vicuna-7b",
+        "instructblip-vicuna-13b",
+        "instructblip-flan-t5-xl",
+        "instructblip-flan-t5-xxl",
+    ]
+    parser.add_argument(
+        "--model_name",
+        default="instructblip-flan-t5-xl",
+        choices=choices,
+        type=str,
+        help="Path to hf config.json of model to convert",
+    )
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument(
+        "--push_to_hub",
+        action="store_true",
+        help="Whether to push the model and processor to the hub after converting",
+    )
+
+    args = parser.parse_args()
+
+    convert_blip2_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/instructblip/processing_instructblip.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 InstructBLIP. Largely copy of Blip2Processor with addition of a tokenizer for the Q-Former.
+"""
+
+import os
+from typing import List, Optional, Union
+
+from ...image_utils import ImageInput
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding, PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from ...utils import TensorType
+from ..auto import AutoTokenizer
+
+
+class InstructBlipProcessor(ProcessorMixin):
+    r"""
+    Constructs an InstructBLIP processor which wraps a BLIP image processor and a LLaMa/T5 tokenizer into a single
+    processor.
+
+    [`InstructBlipProcessor`] offers all the functionalities of [`BlipImageProcessor`] and [`AutoTokenizer`]. See the
+    docstring of [`~BlipProcessor.__call__`] and [`~BlipProcessor.decode`] for more information.
+
+    Args:
+        image_processor (`BlipImageProcessor`):
+            An instance of [`BlipImageProcessor`]. The image processor is a required input.
+        tokenizer (`AutoTokenizer`):
+            An instance of ['PreTrainedTokenizer`]. The tokenizer is a required input.
+        qformer_tokenizer (`AutoTokenizer`):
+            An instance of ['PreTrainedTokenizer`]. The Q-Former tokenizer is a required input.
+    """
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "BlipImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(self, image_processor, tokenizer, qformer_tokenizer):
+        super().__init__(image_processor, tokenizer)
+
+        # add QFormer tokenizer
+        self.qformer_tokenizer = qformer_tokenizer
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        add_special_tokens: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_token_type_ids: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> BatchEncoding:
+        """
+        This method uses [`BlipImageProcessor.__call__`] method to prepare image(s) for the model, and
+        [`BertTokenizerFast.__call__`] to prepare text for the model.
+
+        Please refer to the docstring of the above two methods for more information.
+        """
+        if images is None and text is None:
+            raise ValueError("You have to specify at least images or text.")
+
+        encoding = BatchEncoding()
+
+        if text is not None:
+            text_encoding = self.tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+            encoding.update(text_encoding)
+            qformer_text_encoding = self.qformer_tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+            encoding["qformer_input_ids"] = qformer_text_encoding.pop("input_ids")
+            encoding["qformer_attention_mask"] = qformer_text_encoding.pop("attention_mask")
+
+        if images is not None:
+            image_encoding = self.image_processor(images, return_tensors=return_tensors)
+            encoding.update(image_encoding)
+
+        return encoding
+
+    # Copied from transformers.models.blip.processing_blip.BlipProcessor.batch_decode with BertTokenizerFast->PreTrainedTokenizer
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to PreTrainedTokenizer'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.blip.processing_blip.BlipProcessor.decode with BertTokenizerFast->PreTrainedTokenizer
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to PreTrainedTokenizer'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.blip.processing_blip.BlipProcessor.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))
+
+    # overwrite to save the Q-Former tokenizer in a separate folder
+    def save_pretrained(self, save_directory, **kwargs):
+        if os.path.isfile(save_directory):
+            raise ValueError(f"Provided path ({save_directory}) should be a directory, not a file")
+        os.makedirs(save_directory, exist_ok=True)
+        qformer_tokenizer_path = os.path.join(save_directory, "qformer_tokenizer")
+        self.qformer_tokenizer.save_pretrained(qformer_tokenizer_path)
+        return super().save_pretrained(save_directory, **kwargs)
+
+    # overwrite to load the Q-Former tokenizer from a separate folder
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        qformer_tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="qformer_tokenizer")
+        args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, **kwargs)
+        args.append(qformer_tokenizer)
+        return cls(*args)

src/transformers/utils/dummy_pt_objects.py

@@ -3729,6 +3729,37 @@ class InformerPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+INSTRUCTBLIP_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class InstructBlipForConditionalGeneration(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class InstructBlipPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class InstructBlipQFormerModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class InstructBlipVisionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 JUKEBOX_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/blip_2/test_modeling_blip_2.py

@@ -330,9 +330,7 @@ class Blip2TextModelDecoderOnlyTester:
     def prepare_config_and_inputs(self):
         config = self.get_config()
 
-        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
-            3,
-        )
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(3)
         input_ids[:, -1] = self.eos_token_id  # Eos Token
 
         attention_mask = input_ids.ne(self.pad_token_id)

tests/models/instructblip/test_processor_instructblip.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers.testing_utils import require_vision
+from transformers.utils import is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import (
+        AutoProcessor,
+        BertTokenizerFast,
+        BlipImageProcessor,
+        GPT2Tokenizer,
+        InstructBlipProcessor,
+        PreTrainedTokenizerFast,
+    )
+
+
+@require_vision
+class InstructBlipProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = BlipImageProcessor()
+        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-GPT2Model")
+        qformer_tokenizer = BertTokenizerFast.from_pretrained("hf-internal-testing/tiny-random-bert")
+
+        processor = InstructBlipProcessor(image_processor, tokenizer, qformer_tokenizer)
+
+        processor.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def get_qformer_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).qformer_tokenizer
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = InstructBlipProcessor(
+            tokenizer=self.get_tokenizer(),
+            image_processor=self.get_image_processor(),
+            qformer_tokenizer=self.get_qformer_tokenizer(),
+        )
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
+
+        processor = InstructBlipProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, PreTrainedTokenizerFast)
+
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.image_processor, BlipImageProcessor)
+        self.assertIsInstance(processor.qformer_tokenizer, BertTokenizerFast)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        qformer_tokenizer = self.get_qformer_tokenizer()
+
+        processor = InstructBlipProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
+        )
+
+        image_input = self.prepare_image_inputs()
+
+        input_feat_extract = image_processor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        qformer_tokenizer = self.get_qformer_tokenizer()
+
+        processor = InstructBlipProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
+        )
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tokens = tokenizer(input_str, return_token_type_ids=False)
+        encoded_tokens_qformer = qformer_tokenizer(input_str, return_token_type_ids=False)
+
+        for key in encoded_tokens.keys():
+            self.assertListEqual(encoded_tokens[key], encoded_processor[key])
+
+        for key in encoded_tokens_qformer.keys():
+            self.assertListEqual(encoded_tokens_qformer[key], encoded_processor["qformer_" + key])
+
+    def test_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        qformer_tokenizer = self.get_qformer_tokenizer()
+
+        processor = InstructBlipProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
+        )
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(
+            list(inputs.keys()),
+            ["input_ids", "attention_mask", "qformer_input_ids", "qformer_attention_mask", "pixel_values"],
+        )
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        qformer_tokenizer = self.get_qformer_tokenizer()
+
+        processor = InstructBlipProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
+        )
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)
+
+    def test_model_input_names(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        qformer_tokenizer = self.get_qformer_tokenizer()
+
+        processor = InstructBlipProcessor(
+            tokenizer=tokenizer, image_processor=image_processor, qformer_tokenizer=qformer_tokenizer
+        )
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(
+            list(inputs.keys()),
+            ["input_ids", "attention_mask", "qformer_input_ids", "qformer_attention_mask", "pixel_values"],
+        )

utils/check_repo.py

@@ -57,6 +57,7 @@ PRIVATE_MODELS = [
 # Being in this list is an exception and should **not** be the rule.
 IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     # models to ignore for not tested
+    "InstructBlipQFormerModel",  # Building part of bigger (tested) model.
     "NllbMoeDecoder",
     "NllbMoeEncoder",
     "LlamaDecoder",  # Building part of bigger (tested) model.
@@ -282,6 +283,8 @@ IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     "FlavaMultimodalModel",
     "GPT2DoubleHeadsModel",
     "GPTSw3DoubleHeadsModel",
+    "InstructBlipVisionModel",
+    "InstructBlipQFormerModel",
     "LayoutLMForQuestionAnswering",
     "LukeForMaskedLM",
     "LukeForEntityClassification",