model/blip/modeling_blip_text.py

import math
from typing import Optional, Union

import torch
from torch import Tensor, device, nn
from torch.nn import CrossEntropyLoss

from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache
from transformers.generation import GenerationMixin
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import (
    BaseModelOutputWithPastAndCrossAttentions,
    BaseModelOutputWithPoolingAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
)
from transformers.pytorch_utils import apply_chunking_to_forward
from transformers.utils import logging
from transformer_lens.hook_points import HookPoint, HookedRootModule
from transformers.models.blip.modeling_blip_text import (
    BlipTextEmbeddings, BlipTextSelfOutput, BlipTextIntermediate, 
    BlipTextPooler, BlipTextPreTrainedModel, BlipTextOnlyMLMHead, 
    BlipTextSelfAttention, BlipTextOutput, BlipTextEncoder
)

logger = logging.get_logger(__name__)

    

class HookedBlipTextSelfAttention(BlipTextSelfAttention):
    def __init__(self, config, is_cross_attention, layer_idx=None):
        super().__init__(config, is_cross_attention, layer_idx=layer_idx)
        self.layer_idx = layer_idx
        self.is_cross_attention = is_cross_attention
        
        # Add hooks        
        self.hook_q = HookPoint()
        self.hook_k = HookPoint()
        self.hook_v = HookPoint()
        self.hook_attn_pattern = HookPoint()
        self.hook_attn_out = HookPoint()
        self.hook_resid_pre = HookPoint()
        self.hook_context_layer = HookPoint()

    def save_attn_gradients(self, attn_gradients):
        self.attn_gradients = attn_gradients

    def get_attn_gradients(self):
        return self.attn_gradients

    def save_attention_map(self, attention_map):
        self.attention_map = attention_map

    def get_attention_map(self):
        return self.attention_map

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        cache_position: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor]:
        
        # pre-residual hook
        self.hook_resid_pre.layer_idx = self.layer_idx
        hidden_states = self.hook_resid_pre(hidden_states)
        
        batch_size, seq_length, _ = hidden_states.shape
        query_layer = (
            self.query(hidden_states)
            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
            .transpose(1, 2)
        )
        
        
        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
        is_cross_attention = encoder_hidden_states is not None
        attention_mask = encoder_attention_mask if is_cross_attention else attention_mask

        is_updated = False
        if past_key_values is not None:
            if isinstance(past_key_values, EncoderDecoderCache):
                is_updated = past_key_values.is_updated.get(self.layer_idx)
                if is_cross_attention:
                    # after the first generated id, we can subsequently re-use all key/value_layer from cache
                    curr_past_key_values = past_key_values.cross_attention_cache
                else:
                    curr_past_key_values = past_key_values.self_attention_cache
            else:
                curr_past_key_values = past_key_values

        current_states = encoder_hidden_states if is_cross_attention else hidden_states
        if is_cross_attention and past_key_values is not None and is_updated:
            # reuse k,v, cross_attentions
            key_layer = curr_past_key_values.layers[self.layer_idx].keys
            value_layer = curr_past_key_values.layers[self.layer_idx].values
        else:
            key_layer = (
                self.key(current_states)
                .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
                .transpose(1, 2)
            )
            value_layer = (
                self.value(current_states)
                .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
                .transpose(1, 2)
            )

            if past_key_values is not None:
                # save all key/value_layer to cache to be re-used for fast auto-regressive generation
                cache_position = cache_position if not is_cross_attention else None
                key_layer, value_layer = curr_past_key_values.update(
                    key_layer, value_layer, self.layer_idx, {"cache_position": cache_position}
                )
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                if is_cross_attention and isinstance(past_key_values, EncoderDecoderCache):
                    past_key_values.is_updated[self.layer_idx] = True

        # HOOK q,k,v
        
        # text
        self.hook_q.layer_idx = self.layer_idx
        query_layer = self.hook_q(query_layer)
        
        # image
        self.hook_k.layer_idx = self.layer_idx
        key_layer = self.hook_k(key_layer)
        
        # image
        self.hook_v.layer_idx = self.layer_idx
        value_layer = self.hook_v(value_layer)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in BlipTextModel forward() function)
            attention_scores = attention_scores + attention_mask.to(attention_scores.device)

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)
        
        # HOOK attention pattern
        self.hook_attn_pattern.layer_idx = self.layer_idx
        attention_probs = self.hook_attn_pattern(attention_probs)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs_dropped = self.dropout(attention_probs)

        
        context_layer = torch.matmul(attention_probs_dropped, value_layer)

        # HOOK context layer
        self.hook_context_layer.layer_idx = self.layer_idx
        context_layer = self.hook_context_layer(context_layer)
        
        
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)

        # HOOK attention out
        self.hook_attn_out.layer_idx = self.layer_idx
        self.hook_attn_out(context_layer)

        return context_layer, attention_probs
    
    
class HookedBlipTextAttention(nn.Module):
    def __init__(self, config, is_cross_attention=False, layer_idx=None):
        super().__init__()
        self.self = HookedBlipTextSelfAttention(config, is_cross_attention, layer_idx=layer_idx)
        self.output = BlipTextSelfOutput(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        cache_position: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor]:
        self_outputs = self.self(
            hidden_states,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            cache_position=cache_position,
        )
        attention_output = self.output(self_outputs[0], hidden_states)
        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
        return outputs



class HookedBlipTextOutput(BlipTextOutput):
    def __init__(self, config, layer_idx):
        super().__init__(config)
        self.layer_idx = layer_idx
        self.hook_mlp_out = HookPoint()

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)

        self.hook_mlp_out.layer_idx = self.layer_idx
        self.hook_mlp_out(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states

    
class HookedBlipTextLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_num):
        super().__init__()
        self.config = config
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.layer_num = layer_num
        self.attention = HookedBlipTextAttention(config, layer_idx=layer_num)
        if self.config.is_decoder:
            self.crossattention = HookedBlipTextAttention(
                config, is_cross_attention=self.config.is_decoder, layer_idx=layer_num
            )
        self.intermediate = BlipTextIntermediate(config)
        self.output = HookedBlipTextOutput(config, layer_idx=layer_num)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        cache_position: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor]:
        self_attention_outputs = self.attention(
            hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            past_key_values=past_key_values,
            cache_position=cache_position,
        )
        attention_output = self_attention_outputs[0]
        outputs = self_attention_outputs[1:]

        if encoder_hidden_states is not None:
            cross_attention_outputs = self.crossattention(
                attention_output,
                attention_mask=encoder_attention_mask,
                encoder_hidden_states=encoder_hidden_states,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                cache_position=cache_position,
            )
            attention_output = cross_attention_outputs[0]
            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
        layer_output = apply_chunking_to_forward(
            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
        )
        return (layer_output,) + outputs

    def feed_forward_chunk(self, attention_output):
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output

        
class HookedBlipTextEncoder(BlipTextEncoder, HookedRootModule):
    def __init__(self, config):
        HookedRootModule.__init__(self)
        self.config = config
        self.layer = nn.ModuleList([HookedBlipTextLayer(config, i) for i in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Cache] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = False,
        output_hidden_states: Optional[bool] = False,
        return_dict: Optional[bool] = True,
        cache_position: Optional[torch.Tensor] = None,
    ) -> Union[tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        if use_cache:
            # The model acts as encoder decoder but is not an encoder decoder. So we cast all cache objects to
            # `EncoderDecoderCache` type assuming that the incoming cache is from `self_attention`
            if isinstance(past_key_values, DynamicCache):
                past_key_values = EncoderDecoderCache(past_key_values, DynamicCache(config=self.config))
            elif past_key_values is None:
                past_key_values = EncoderDecoderCache(
                    DynamicCache(config=self.config), DynamicCache(config=self.config)
                )

        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None
        all_cross_attentions = () if output_attentions and encoder_hidden_states is not None else None

        for i in range(self.config.num_hidden_layers):
            layer_module = self.layer[i]
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            layer_outputs = layer_module(
                hidden_states,
                attention_mask,
                encoder_hidden_states,
                encoder_attention_mask,
                past_key_values,
                output_attentions,
                cache_position,
            )

            hidden_states = layer_outputs[0]
            if output_attentions:
                all_self_attentions = all_self_attentions + (layer_outputs[1],)
                if encoder_hidden_states is not None:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [
                    hidden_states,
                    past_key_values,
                    all_hidden_states,
                    all_self_attentions,
                    all_cross_attentions,
                ]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
            cross_attentions=all_cross_attentions,
        )


# Adapted from https://github.com/salesforce/BLIP/blob/3a29b7410476bf5f2ba0955827390eb6ea1f4f9d/models/med.py#L571
class BlipTextModel(BlipTextPreTrainedModel):
    """
    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
    cross-attention is added between the self-attention layers, following the architecture described in [Attention is
    all you need](https://huggingface.co/papers/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. argument and `is_decoder` set to `True`; an
    `encoder_hidden_states` is then expected as an input to the forward pass.
    """

    def __init__(self, config, add_pooling_layer=True):
        super().__init__(config)
        self.config = config

        self.embeddings = BlipTextEmbeddings(config)
        self.encoder = HookedBlipTextEncoder(config)
        self.pooler = BlipTextPooler(config) if add_pooling_layer else None
        
        self.hook_text_embeddings = HookPoint()

        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    def get_extended_attention_mask(
        self, attention_mask: Tensor, input_shape: tuple[int], device: device, is_decoder: bool
    ) -> Tensor:
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

        Arguments:
            attention_mask (`torch.Tensor`):
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
            input_shape (`tuple[int]`):
                The shape of the input to the model.
            device (`torch.device`):
                The device of the input to the model.

        Returns:
            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
        """
        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        if attention_mask.dim() == 3:
            extended_attention_mask = attention_mask[:, None, :, :]
        elif attention_mask.dim() == 2:
            # Provided a padding mask of dimensions [batch_size, seq_length]
            # - if the model is a decoder, apply a causal mask in addition to the padding mask
            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
            if is_decoder:
                batch_size, seq_length = input_shape

                seq_ids = torch.arange(seq_length, device=device)
                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
                # in case past_key_values are used we need to add a prefix ones mask to the causal mask
                causal_mask = causal_mask.to(attention_mask.dtype)

                if causal_mask.shape[1] < attention_mask.shape[1]:
                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
                    causal_mask = torch.cat(
                        [
                            torch.ones(
                                (batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype
                            ),
                            causal_mask,
                        ],
                        axis=-1,
                    )

                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
            else:
                extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
            )

        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
        # masked positions, this operation will create a tensor which is 0.0 for
        # positions we want to attend and -10000.0 for masked positions.
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
        return extended_attention_mask

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        encoder_embeds: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        is_decoder: Optional[bool] = False,
        cache_position: Optional[torch.Tensor] = None,
    ) -> Union[tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
        r"""
        encoder_hidden_states  (`torch.FloatTensor`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
            the model is configured as a decoder.
        encoder_attention_mask (`torch.FloatTensor`, *optional*):
            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.
        past_key_values (`Cache`, *optional*):
            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up 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)`.
        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 = 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

        if is_decoder:
            use_cache = use_cache if use_cache is not None else self.config.use_cache
        else:
            use_cache = False

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
            batch_size, seq_length = input_shape
            device = input_ids.device
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size, seq_length = input_shape
            device = inputs_embeds.device
        elif encoder_embeds is not None:
            input_shape = encoder_embeds.size()[:-1]
            batch_size, seq_length = input_shape
            device = encoder_embeds.device
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds")

        past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()

        if attention_mask is None:
            attention_mask = torch.ones((batch_size, seq_length + past_key_values_length)).to(device)

        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
            attention_mask, input_shape, device, is_decoder
        )

        # If a 2D or 3D attention mask is provided for the cross-attention
        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if encoder_hidden_states is not None:
            if isinstance(encoder_hidden_states, list):
                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
            else:
                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)

            if isinstance(encoder_attention_mask, list):
                encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
            elif encoder_attention_mask is None:
                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
            else:
                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
        else:
            encoder_extended_attention_mask = None

        if encoder_embeds is None:
            embedding_output = self.embeddings(
                input_ids=input_ids,
                position_ids=position_ids,
                inputs_embeds=inputs_embeds,
                past_key_values_length=past_key_values_length,
            )
        else:
            embedding_output = encoder_embeds
        
        # HOOK text embeddings
        embedding_output = self.hook_text_embeddings(embedding_output)

        encoder_outputs = self.encoder(
            embedding_output,
            attention_mask=extended_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_extended_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )
        sequence_output = encoder_outputs[0]
        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None

        if not return_dict:
            return (sequence_output, pooled_output) + encoder_outputs[1:]

        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            past_key_values=encoder_outputs.past_key_values,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
            cross_attentions=encoder_outputs.cross_attentions,
        )


# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L811
class BlipTextLMHeadModel(BlipTextPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]

    def __init__(self, config):
        super().__init__(config)

        self.bert = BlipTextModel(config, add_pooling_layer=False)
        self.cls = BlipTextOnlyMLMHead(config)
        self.label_smoothing = config.label_smoothing

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

    def set_input_embeddings(self, new_embeddings):
        self.bert.set_input_embeddings(new_embeddings)

    def get_output_embeddings(self):
        return self.cls.predictions.decoder

    def set_output_embeddings(self, new_embeddings):
        self.cls.predictions.decoder = new_embeddings
        self.cls.predictions.bias = new_embeddings.bias

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        return_logits: Optional[bool] = False,
        is_decoder: Optional[bool] = True,
        reduction: Optional[str] = "mean",
        cache_position: Optional[torch.Tensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
    ) -> Union[tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
        r"""
        encoder_hidden_states (`torch.FloatTensor`, *optional*): Sequence of
            hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is
            configured as a decoder.
        encoder_attention_mask (`torch.FloatTensor`, *optional*):
            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.
        labels (`torch.LongTensor`, *optional*):
            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
            `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
            ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`
        past_key_values (`Cache`, *optional*):
            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up 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)`.
        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`).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if labels is not None:
            use_cache = False

        outputs = self.bert(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            is_decoder=is_decoder,
            cache_position=cache_position,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        prediction_scores = self.cls(hidden_states[:, slice_indices, :])

        if return_logits:
            return prediction_scores[:, :-1, :].contiguous()

        lm_loss = None
        if labels is not None:
            # we are doing next-token prediction; shift prediction scores and input ids by one
            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
            labels = labels[:, 1:].contiguous().to(shifted_prediction_scores.device)
            loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=self.label_smoothing)
            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
            if reduction == "none":
                lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)

        if not return_dict:
            output = (prediction_scores,) + outputs[2:]
            return ((lm_loss,) + output) if lm_loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=lm_loss,
            logits=prediction_scores,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            cross_attentions=outputs.cross_attentions,
        )

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, **model_kwargs):
        # Overwrite -- hardcoded key return (`is_decoder=True`)

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            **model_kwargs,
        )
        model_inputs["is_decoder"] = True

        return model_inputs


__all__ = ["BlipTextModel", "BlipTextLMHeadModel", "BlipTextPreTrainedModel"]