venv/lib/python3.12/site-packages/transformers/models/paligemma/modeling_paligemma.py

# coding=utf-8
# 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 dataclasses import dataclass
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn

from ...cache_utils import Cache
from ...modeling_utils import PreTrainedModel
from ...utils import (
    ModelOutput,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    logging,
    replace_return_docstrings,
)
from .configuration_paligemma import PaliGemmaConfig


if is_flash_attn_2_available():
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa

from ..auto import AutoModel, AutoModelForCausalLM


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "PaliGemmaConfig"


@dataclass
class PaliGemmaCausalLMOutputWithPast(ModelOutput):
    """
    Base class for PaliGemmacausal language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
            `past_key_values` input) to speed up sequential decoding.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
            Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
            sequence_length, hidden_size)`.

            image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
    """

    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    image_hidden_states: Optional[Tuple[torch.FloatTensor]] = 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


PALIGEMMA_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`PaliGemmaConfig`] or [`PaliGemmaVisionConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
    PALIGEMMA_START_DOCSTRING,
)
class PaliGemmaPreTrainedModel(PreTrainedModel):
    config_class = PaliGemmaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["PaliGemmaMultiModalProjector"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = False
    _supports_sdpa = True
    _supports_cache_class = True

    def _init_weights(self, module):
        # important: this ported version of PaliGemmaisn't meant for training from scratch - only
        # inference and fine-tuning
        std = (
            self.config.initializer_range
            if hasattr(self.config, "initializer_range")
            else self.config.text_config.initializer_range
        )

        if hasattr(module, "class_embedding"):
            module.class_embedding.data.normal_(mean=0.0, std=std)

        if isinstance(module, (nn.Linear, nn.Conv2d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

    @property
    def _supports_sdpa(self):
        """
        Retrieve language_model's attribute to check whether the model supports
        SDPA or not.
        """
        return self.language_model._supports_sdpa


PALIGEMMA_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
            The tensors corresponding to the input images. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`SiglipImageProcessor.__call__`] for details ([]`PaliGemmaProcessor`] uses
            [`SiglipImageProcessor`] for processing images).
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


@add_start_docstrings(
    """The PALIGEMMA model which consists of a vision backbone and a language model.""",
    PALIGEMMA_START_DOCSTRING,
)
class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
    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
        self._attn_implementation = config._attn_implementation

        language_model = AutoModelForCausalLM.from_config(
            config=config.text_config, attn_implementation=self._attn_implementation
        )

        if language_model._tied_weights_keys is not None:
            self._tied_weights_keys = [f"language_model.{k}" for k in language_model._tied_weights_keys]
        self.language_model = language_model

        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
        self.post_init()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.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.LlavaForConditionalGeneration.set_input_embeddings with Llava->PaliGemma
    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

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

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_output_embeddings with Llava->PaliGemma
    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_decoder with Llava->PaliGemma
    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

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

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

    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
        # TODO: config.vocab_size is deprecated and will be removed in v4.43.
        # `resize_token_embeddings` should work from `modeling_utils.py``
        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
        self.config.text_config.vocab_size = model_embeds.num_embeddings
        self.config.vocab_size = model_embeds.num_embeddings
        self.vocab_size = model_embeds.num_embeddings
        return model_embeds

    def _merge_input_ids_with_image_features(
        self, image_features, inputs_embeds, input_ids, attention_mask, labels, token_type_ids, cache_position
    ):
        _, _, embed_dim = image_features.shape
        batch_size, sequence_length = input_ids.shape
        dtype, device = inputs_embeds.dtype, inputs_embeds.device
        min_dtype = torch.finfo(dtype).min

        scaled_image_features = image_features / (self.config.hidden_size**0.5)
        final_embedding = torch.zeros(
            batch_size, sequence_length, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
        )

        text_mask = (input_ids != self.config.image_token_index) & (input_ids != self.pad_token_id)
        image_mask = input_ids == self.config.image_token_index
        pad_mask = input_ids == self.pad_token_id

        # expand masks to match embedding dimension
        text_mask_expanded = text_mask.unsqueeze(-1).expand(-1, -1, embed_dim).to(inputs_embeds.device)
        pad_mask_expanded = pad_mask.unsqueeze(-1).expand(-1, -1, embed_dim).to(inputs_embeds.device)
        # insert padding and text token embeddings
        final_embedding = torch.where(text_mask_expanded, inputs_embeds, final_embedding)
        final_embedding = torch.where(pad_mask_expanded, torch.zeros_like(final_embedding), final_embedding)
        # insert image embeddings - the image mask is always less or equal to the sentence in length
        final_embedding = final_embedding.masked_scatter(
            image_mask.unsqueeze(-1).expand_as(final_embedding).to(device=final_embedding.device),
            scaled_image_features.to(device=final_embedding.device, dtype=final_embedding.dtype),
        )
        final_embedding = torch.where(pad_mask_expanded, torch.zeros_like(final_embedding), final_embedding)
        if attention_mask is not None:
            position_ids = (attention_mask.cumsum(-1)).masked_fill_((attention_mask == 0), 1)
        else:
            position_ids = None

        if token_type_ids is not None and labels is not None:
            # we are training thus we need to create a full mask on the image + prefix but causal on suffix
            target_length = cache_position[-1] + 1
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(inputs_embeds.shape[0], 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                # unmask the prefill
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    token_type_ids[:, None, None, :].to(causal_mask.device) == 0, 0
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

            final_labels = torch.full(
                (batch_size, sequence_length), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
            )
            final_labels = torch.where(input_ids != self.pad_token_id, labels, final_labels)
        else:
            causal_mask = attention_mask.unsqueeze(1).unsqueeze(2) * attention_mask.unsqueeze(1).unsqueeze(-1)
            # invert causal mask
            causal_mask = torch.where(causal_mask == 0, min_dtype, 0)
            causal_mask = causal_mask.to(dtype).expand(-1, self.config.text_config.num_key_value_heads, -1, -1)
            final_labels = None

        return final_embedding, causal_mask, final_labels, position_ids

    @add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, PaliGemmaCausalLMOutputWithPast]:
        r"""
        Args:
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Returns:

        Example:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, PaliGemmaForConditionalGeneration

        >>> model = PaliGemmaForConditionalGeneration.from_pretrained("google/PaliGemma-test-224px-hf")
        >>> processor = AutoProcessor.from_pretrained("google/PaliGemma-test-224px-hf")

        >>> prompt = "answer en Where is the cow standing?"
        >>> url = "https://huggingface.co/gv-hf/PaliGemma-test-224px-hf/resolve/main/cow_beach_1.png"
        >>> image = Image.open(requests.get(url, stream=True).raw)

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

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_length=30)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "answer en Where is the cow standing?\nbeach"
        ```"""

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
            )

        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

        # the attention mask is turned 4d after, we keep track of the original one
        input_attention_mask = attention_mask

        if inputs_embeds is None:
            # 1. Extra the input embeddings
            inputs_embeds = self.get_input_embeddings()(input_ids)

            # 2. Merge text and images
            if pixel_values is not None and input_ids.shape[1] != 1:
                image_outputs = self.vision_tower(pixel_values.to(inputs_embeds.dtype))
                selected_image_feature = image_outputs.last_hidden_state
                image_features = self.multi_modal_projector(selected_image_feature)

                if cache_position is None:
                    cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device)
                inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
                    image_features, inputs_embeds, input_ids, attention_mask, labels, token_type_ids, cache_position
                )

            else:
                # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
                # generation with cache
                if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
                    # Retrieve the first layer to inspect the logits and mask out the hidden states
                    # that are set to 0
                    # TODO @molbap this will only work for dynamic cache.
                    first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

                    # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
                    batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

                    # Get the target length
                    target_seqlen = cache_position[-1] + 1
                    extended_attention_mask = torch.ones(
                        (attention_mask.shape[0], target_seqlen - attention_mask.shape[1] + 1),
                        dtype=attention_mask.dtype,
                        device=attention_mask.device,
                    )
                    # Filter out only the tokens that can be un-attended, this can happen
                    # if one uses PaliGemma+ Fused modules where the cache on the
                    # first iteration is already big enough, or if one passes custom cache
                    valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
                    new_batch_index = batch_index[valid_indices]
                    new_non_attended_tokens = non_attended_tokens[valid_indices]

                    # Zero-out the places where we don't need to attend
                    extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0

                    attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)
                    position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1

        attention_mask = attention_mask.to(inputs_embeds.dtype)
        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        logits = outputs.logits
        logits = logits.float()
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :]
            shift_labels = labels[..., 1:]
            if input_attention_mask is not None:
                # we use the input attention mask to shift the logits and labels, because it is 2D.
                shift_attention_mask = input_attention_mask[..., 1:]
                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
            else:
                shift_logits = shift_logits.contiguous()
                shift_labels = shift_labels.contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()

            flat_logits = shift_logits.view(-1, self.config.vocab_size)
            flat_labels = shift_labels.view(-1).to(shift_logits.device)
            loss = loss_fct(flat_logits, flat_labels)
        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

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

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        pixel_values=None,
        attention_mask=None,
        token_type_ids=None,
        use_cache=True,
        **kwargs,
    ):
        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        if past_key_values is not None:
            if inputs_embeds is not None:  # Exception 1
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and cache_position[0] == 0:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "cache_position": cache_position,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "token_type_ids": token_type_ids,
            }
        )
        return model_inputs