image-match/python/py/Lib/site-packages/transformers/models/paligemma/modeling_paligemma.py

# Copyright 2024 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 PaliGemmamodel."""

from collections.abc import Callable
from dataclasses import dataclass

import torch
from torch import nn

from ...cache_utils import Cache
from ...configuration_utils import PreTrainedConfig
from ...generation import GenerationMixin
from ...masking_utils import create_masks_for_generate
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling
from ...modeling_utils import PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    ModelOutput,
    TransformersKwargs,
    auto_docstring,
    can_return_tuple,
    logging,
    torch_compilable_check,
)
from ...utils.deprecation import deprecate_kwarg
from ..auto import AutoModel
from .configuration_paligemma import PaliGemmaConfig


logger = logging.get_logger(__name__)


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Paligemma outputs, with hidden states and attentions.
    """
)
class PaligemmaModelOutputWithPast(BaseModelOutputWithPast):
    r"""
    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 PaliGemma causal language model (or autoregressive) outputs.
    """
)
class PaliGemmaCausalLMOutputWithPast(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.text_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 after projecting 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 PaliGemmaMultiModalProjector(nn.Module):
    def __init__(self, config: PaliGemmaConfig):
        super().__init__()
        self.linear = nn.Linear(config.vision_config.hidden_size, config.vision_config.projection_dim, bias=True)

    def forward(self, image_features):
        hidden_states = self.linear(image_features)

        return hidden_states


def token_type_ids_mask_function(group_ids: torch.Tensor) -> Callable:
    """
    This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths,
    not start and end indices.
    Args:
        group_ids (`torch.Tensor`):
            A tensor of shape `(bs, len)` assigning each token to a vision group. Tokens with the same group
            come from the same input image. Text is denoted by `-1`.
    """

    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
        seq_length = group_ids.shape[-1]

        # clamp indices because with static cache they can go beyond `group_ids.shape[-1]`
        q_idx_clamped = q_idx.clamp(max=seq_length - 1)
        kv_idx_clamped = kv_idx.clamp(max=seq_length - 1)

        # Unmask if the q and kv come from same group which is not -1 (i.e. non-text)
        q_group = group_ids[batch_idx, q_idx_clamped]
        kv_group = group_ids[batch_idx, kv_idx_clamped]
        q_group = torch.where(q_idx < seq_length, q_group, -1)
        kv_group = torch.where(kv_idx < seq_length, kv_group, -1)
        return (q_group == kv_group) & (q_group >= 0)

    return inner_mask


@deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds")
def create_causal_mask_mapping(
    config: PreTrainedConfig,
    inputs_embeds: torch.Tensor,
    attention_mask: torch.Tensor | None,
    past_key_values: Cache | None,
    position_ids: torch.Tensor | None,
    token_type_ids: torch.Tensor | None = None,
    pixel_values: torch.FloatTensor | None = None,
    is_training: bool | None = False,
    is_first_iteration: bool | None = None,
    **kwargs,
) -> dict:
    """
    Overwrites the base `create_masks_for_generate` with `token_type_ids` masking to create the causal mask mapping
    for all kinds of forward passes. Paligemma uses a bidirectional mask on the prompt tokens.

    Uses `pixel_values` as an optional input to disambiguate edge cases.
    """
    if is_training and token_type_ids is None:
        raise ValueError("`token_type_ids` is required as a model input when training")

    mask_kwargs = {
        "config": config.get_text_config(),
        "inputs_embeds": inputs_embeds,
        "attention_mask": attention_mask,
        "past_key_values": past_key_values,
        "position_ids": position_ids,
    }
    # Infer if prefill or decoding stage, if the flag isn't passed. This happens only when the mask is constructed
    # from `forward` call. If users run a `forward` call, we have no option to infer `is_first_iteration` because users may be
    # running generation with custom loop. Thus we need to infer it in a `non-perfect` way
    # NOTE: Determining prefill in that case requires checking data values, which is not compile-compatible.
    is_first_iteration = (
        is_first_iteration
        if is_first_iteration
        else (past_key_values is None or not past_key_values.is_initialized or pixel_values is not None)
    )

    if is_first_iteration or not kwargs.get("use_cache", True):
        if token_type_ids is not None:
            # The logic bellow was originally written for Gemma3, where `token_type_ids` is reversed. Let's reverse
            # it to then use exactly the same logic.
            token_type_ids = 1 - token_type_ids
        else:
            logger.warning_once(
                "It is a prefill stage but The `token_type_ids` is not provided. We recommend "
                "passing `token_type_ids` to the model to prevent bad attention masking."
            )
            # NOTE: this branch can't be reached when training because `token_type_ids` is required as a model input.
            token_type_ids = torch.ones_like(inputs_embeds)[:, :, 0]

    # Logic originally copied from Gemma3. It holds up for Paligemma as well because Paligemma assumes up to one image
    # per prompt AND we reverse `token_type_ids` above. Gemma3 uses a bidirectional mask for images, tagged through
    # `token_type_ids` 1s.
    if token_type_ids is not None and is_first_iteration:
        # We need to pass an additional mask function to account for token type ids, and it needs to be an `or` (to
        # undo the causal masking)

        # First find where a new image block starts: 1 if image and previous not image
        # The images cannot attend to future images, but can attend to all prev images and to itself bidirectionally
        is_image = (token_type_ids == 1).to(inputs_embeds.device)
        is_previous_image = nn.functional.pad(is_image, (1, 0), value=0)[:, :-1]
        new_image_start = is_image & ~is_previous_image
        group_ids = torch.cumsum(new_image_start.int(), dim=1) - 1
        group_ids = torch.where(is_image, group_ids, torch.full_like(token_type_ids, -1))
        mask_kwargs["or_mask_function"] = token_type_ids_mask_function(group_ids)

    return create_masks_for_generate(**mask_kwargs)


@auto_docstring
class PaliGemmaPreTrainedModel(PreTrainedModel):
    config: PaliGemmaConfig
    base_model_prefix = "model"
    input_modalities = ("image", "text")
    supports_gradient_checkpointing = True
    _no_split_modules = ["PaliGemmaMultiModalProjector"]
    _skip_keys_device_placement = "past_key_values"
    _can_compile_fullgraph = False
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_attention_backend = True


@auto_docstring(
    custom_intro="""
    The Base Paligemma model which consists of a vision backbone and a language model without language modeling head.,
    """
)
class PaliGemmaModel(PaliGemmaPreTrainedModel):
    # we are filtering the logits/labels so we shouldn't divide the loss based on num_items_in_batch
    accepts_loss_kwargs = False

    def __init__(self, config: PaliGemmaConfig):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config=config.vision_config)
        self.multi_modal_projector = PaliGemmaMultiModalProjector(config)
        self.vocab_size = config.text_config.vocab_size

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

        self.text_config_dtype = self.config.get_text_config().dtype or self.dtype
        self.post_init()

    # Copied from transformers.models.llava.modeling_llava.LlavaModel.get_input_embeddings with Llava->PaliGemma
    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    # Copied from transformers.models.llava.modeling_llava.LlavaModel.set_input_embeddings with Llava->PaliGemma
    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, **kwargs: Unpack[TransformersKwargs]
    ) -> tuple | BaseModelOutputWithPooling:
        image_outputs = self.vision_tower(pixel_values, **kwargs)
        selected_image_feature = image_outputs.last_hidden_state
        image_features = self.multi_modal_projector(selected_image_feature)
        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

    @can_return_tuple
    @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,
        token_type_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple | PaligemmaModelOutputWithPast:
        r"""
        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.text_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.text_config.vocab_size]`.

        Example:

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

        >>> model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma2-3b-mix-224")
        >>> processor = AutoProcessor.from_pretrained("google/paligemma2-3b-mix-224")

        >>> prompt = "Where is the cat standing?"
        >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs,)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Where is the cat standing?\nsnow"
        ```"""

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

        # Replace image id with PAD if the image token if OOV, to avoid index-errors
        if input_ids is not None and self.config.image_token_id >= self.vocab_size:
            special_image_mask = input_ids == self.config.image_token_id
            llm_input_ids = input_ids.clone()
            llm_input_ids[special_image_mask] = 0
        else:
            llm_input_ids = input_ids

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

        if position_ids is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
            position_ids = position_ids.unsqueeze(0) + 1  # Paligemma positions are 1-indexed

        # Merge text and images
        if pixel_values is not None:
            image_features = self.get_image_features(pixel_values).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)

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            causal_mask_mapping = create_causal_mask_mapping(
                self.config,
                inputs_embeds,
                attention_mask,
                past_key_values,
                position_ids,
                token_type_ids,
                pixel_values,
                is_training=self.training,
            )

        outputs = self.language_model(
            attention_mask=causal_mask_mapping,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        return PaligemmaModelOutputWithPast(
            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,
        )


@auto_docstring(
    custom_intro="""
    The Base Paligemma model which consists of a vision backbone and a language model without language modeling head.,
    """
)
class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}

    def __init__(self, config: PaliGemmaConfig):
        super().__init__(config)
        self.model = PaliGemmaModel(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)

    @auto_docstring
    def get_image_features(self, pixel_values: torch.FloatTensor, **kwargs: Unpack[TransformersKwargs]):
        return self.model.get_image_features(pixel_values, **kwargs)

    @can_return_tuple
    @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,
        token_type_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | PaliGemmaCausalLMOutputWithPast:
        r"""
        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.text_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.text_config.vocab_size]`.

        Example:

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

        >>> model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma2-3b-mix-224")
        >>> processor = AutoProcessor.from_pretrained("google/paligemma2-3b-mix-224")

        >>> prompt = "Where is the cat standing?"
        >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs,)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Where is the cat standing?\nsnow"
        ```"""
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            token_type_ids=token_type_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            labels=labels,
            **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, **kwargs
            )

        return PaliGemmaCausalLMOutputWithPast(
            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,
        position_ids=None,
        pixel_values=None,
        attention_mask=None,
        token_type_ids=None,
        use_cache=True,
        logits_to_keep=None,
        labels=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- custom `position_ids` and `pixel_values` handling
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            use_cache=use_cache,
            logits_to_keep=logits_to_keep,
            token_type_ids=token_type_ids,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        # position_ids in Paligemma are 1-indexed
        if model_inputs.get("position_ids") is not None:
            # NOTE: we need this op out-of-place, otherwise it modifies the `model_kwargs` dict used in `generate` in-place!
            model_inputs["position_ids"] = model_inputs["position_ids"] + 1

        # Pixel values are used only in the first iteration if available
        # In subsequent 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). NOTE: use_cache=False needs pixel_values always
        if is_first_iteration or not use_cache:
            model_inputs["pixel_values"] = pixel_values

        return model_inputs

    @staticmethod
    @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds")
    def create_masks_for_generate(
        config: PreTrainedConfig,
        inputs_embeds: torch.Tensor,
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None,
        position_ids: torch.Tensor | None,
        token_type_ids: torch.Tensor | None = None,
        is_first_iteration: bool | None = False,
        **kwargs,
    ) -> dict:
        # Uses the overwritten `create_masks_for_generate` with `token_type_ids` masking
        return create_causal_mask_mapping(
            config,
            inputs_embeds,
            attention_mask,
            past_key_values,
            position_ids,
            token_type_ids,
            is_first_iteration=is_first_iteration,
            **{k: v for k, v in kwargs.items() if k != "pixel_values"},
        )


__all__ = ["PaliGemmaForConditionalGeneration", "PaliGemmaPreTrainedModel", "PaliGemmaModel"]