webapp/seed/Lib/site-packages/transformers/models/vipllava/modeling_vipllava.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/vipllava/modular_vipllava.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_vipllava.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# 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.

from dataclasses import dataclass

import torch
from torch import nn

from ...activations import ACT2FN
from ...cache_utils import Cache
from ...generation import GenerationMixin
from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, ModelOutput
from ...modeling_utils import PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check
from ...utils.generic import check_model_inputs
from ..auto import AutoModel
from .configuration_vipllava import VipLlavaConfig


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for VipLlava outputs, with hidden states and attentions.
    """
)
class VipLlavaModelOutputWithPast(BaseModelOutputWithPast):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        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.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    image_hidden_states: torch.FloatTensor | None = None


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for VipLlava causal language model (or autoregressive) outputs.
    """
)
class VipLlavaCausalLMOutputWithPast(ModelOutput):
    r"""
    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 (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        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.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: torch.FloatTensor | None = None


class VipLlavaMultiModalProjector(nn.Module):
    def __init__(self, config: VipLlavaConfig):
        super().__init__()
        num_feature_layers = 1 if isinstance(config.vision_feature_layers, int) else len(config.vision_feature_layers)
        self.projector_layernorm = nn.LayerNorm(
            num_feature_layers * config.vision_config.hidden_size, eps=config.projector_layernorm_eps
        )

        self.linear_1 = nn.Linear(
            num_feature_layers * 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, hidden_states):
        hidden_states = self.projector_layernorm(hidden_states)
        hidden_states = self.linear_1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


@auto_docstring
class VipLlavaPreTrainedModel(PreTrainedModel):
    config: VipLlavaConfig
    base_model_prefix = "model"
    input_modalities = ("image", "text")
    supports_gradient_checkpointing = True
    _skip_keys_device_placement = "past_key_values"

    _supports_flash_attn = True
    _supports_sdpa = True

    _can_compile_fullgraph = True
    _supports_flex_attn = True
    _supports_attention_backend = True


@auto_docstring(
    custom_intro="""
    The VipLlava model which consists of a vision backbone and a language model, without a language modeling head.
    """
)
class VipLlavaModel(VipLlavaPreTrainedModel):
    _checkpoint_conversion_mapping = {
        r"^language_model.model": "language_model",
    }

    def __init__(self, config: VipLlavaConfig):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)

        self.multi_modal_projector = VipLlavaMultiModalProjector(config)
        self.language_model = AutoModel.from_config(config.text_config)
        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)

    @can_return_tuple
    @auto_docstring(
        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
    )
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        vision_feature_layers: int | list[int] | None = None,
        output_hidden_states: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`):
            The tensors corresponding to the input images.
        vision_feature_layers (`Union[int, list[int]]`, *optional*):
            The vision feature layer, or the list of indexes of the layers to select
            the vision feature.
        """
        vision_feature_layers = (
            vision_feature_layers if vision_feature_layers is not None else self.config.vision_feature_layers
        )
        image_outputs = self.vision_tower(
            pixel_values,
            output_hidden_states=True,  # Ignore arg on purpose
            return_dict=True,
            **kwargs,
        )

        # If multiple feature layers are provided (which is usually the case)
        # then the image features are concatenated after the CLS is removed.
        if isinstance(vision_feature_layers, int):
            image_features = image_outputs.hidden_states[vision_feature_layers][:, 1:]
        else:
            # Usually, we select the features from index 1: the layers -2, -5, -8, -11 and 6
            image_features = [image_outputs.hidden_states[index][:, 1:] for index in vision_feature_layers]
            image_features = torch.cat(image_features, dim=-1)
        image_features = self.multi_modal_projector(image_features)
        image_outputs.pooler_output = image_features

        return image_outputs

    def get_placeholder_mask(
        self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id

        n_image_tokens = special_image_mask.sum()
        n_image_features = image_features.shape[0] * image_features.shape[1]
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        torch_compilable_check(
            inputs_embeds[special_image_mask].numel() == image_features.numel(),
            f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {n_image_features}",
        )
        return special_image_mask

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        vision_feature_layers: int | list[int] | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        **lm_kwargs,
    ) -> tuple | VipLlavaModelOutputWithPast:
        r"""
        vision_feature_layers (`Union[int, list[int]]`, *optional*):
            The vision feature layer, or the list of indexes of the layers to select
            the vision feature.
        """
        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_layers = (
            vision_feature_layers if vision_feature_layers is not None else self.config.vision_feature_layers
        )

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if pixel_values is not None:
            image_features = self.get_image_features(
                pixel_values=pixel_values, vision_feature_layers=vision_feature_layers, return_dict=True
            ).pooler_output
            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
            )
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        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=True,
            cache_position=cache_position,
            **lm_kwargs,
        )

        output = VipLlavaModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )
        return output if return_dict else output.to_tuple()


@auto_docstring(
    custom_intro="""
    The VIPLLAVA model which consists of a vision backbone and a language model.
    """
)
class VipLlavaForConditionalGeneration(VipLlavaPreTrainedModel, GenerationMixin):
    _checkpoint_conversion_mapping = {
        r"^language_model.model": "model.language_model",
        r"^vision_tower": "model.vision_tower",
        r"^multi_modal_projector": "model.multi_modal_projector",
        r"^language_model.lm_head": "lm_head",
    }
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}

    def __init__(self, config: VipLlavaConfig):
        super().__init__(config)
        self.model = VipLlavaModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.post_init()

    def get_input_embeddings(self):
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

    def get_output_embeddings(self) -> nn.Module:
        return self.lm_head

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        vision_feature_layers: int | list[int] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`):
            The tensors corresponding to the input images.
        vision_feature_layers (`Union[int, list[int]]`, *optional*):
            The vision feature layer, or the list of indexes of the layers to select
            the vision feature.
        """
        return self.model.get_image_features(
            pixel_values=pixel_values, vision_feature_layers=vision_feature_layers, **kwargs
        )

    @check_model_inputs(tie_last_hidden_states=False)
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        vision_feature_layers: int | list[int] | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **lm_kwargs,
    ) -> tuple | VipLlavaCausalLMOutputWithPast:
        r"""
        vision_feature_layers (`Union[int, list[int]]`, *optional*):
            The vision feature layer, or the list of indexes of the layers to select
            the vision feature.
        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]`.

        Example:

        ```python
        >>> import torch
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, VipLlavaForConditionalGeneration

        >>> model = VipLlavaForConditionalGeneration.from_pretrained("llava-hf/vip-llava-7b-hf", device_map="auto", dtype=torch.float16)
        >>> processor = AutoProcessor.from_pretrained("llava-hf/vip-llava-7b-hf")

        >>> prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.###Human: <image>\n{}###Assistant:"
        >>> question = "Can you please describe this image?"
        >>> prompt = prompt.format(question)
        >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-neg.png"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> inputs = processor(text=text, images=image, return_tensors="pt").to(0, torch.float16)

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_new_tokens=20)
        >>> processor.decode(generate_ids[0][len(inputs["input_ids"][0]):], skip_special_tokens=True)
        The image features a brown and white cat sitting on a green surface, with a red ball in its
        ```"""

        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_layers = (
            vision_feature_layers if vision_feature_layers is not None else self.config.vision_feature_layers
        )

        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            vision_feature_layers=vision_feature_layers,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **lm_kwargs,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size)

        return VipLlavaCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        attention_mask=None,
        cache_position=None,
        logits_to_keep=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            logits_to_keep=logits_to_keep,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        if is_first_iteration or not kwargs.get("use_cache", True):
            # Pixel values are used only in the first iteration if available
            # In subsquent iterations, they are already merged with text and cached
            # NOTE: first iteration doesn't have to be prefill, it can be the first
            # iteration with a question and cached system prompt (continue generate from cache)
            model_inputs["pixel_values"] = pixel_values

        return model_inputs


__all__ = ["VipLlavaModel", "VipLlavaForConditionalGeneration", "VipLlavaPreTrainedModel"]