m4/models/vopt/modeling_vopt.py

# coding=utf-8
# Copyright 2022 The Fairseq Authors and 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 OPT model."""
import random
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss
from transformers.activations import ACT2FN
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.utils import (
    ContextManagers,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)

from m4.models import DecoupledEmbedding, DecoupledLinear
from m4.models.common import (
    expand_inputs_for_generation,
    prepare_inputs_for_generation,
    update_model_kwargs_for_generation,
)
from m4.models.custom_modules import VLOOMPreTrainedModelBase
from m4.models.perceiver.perceiver import PerceiverResampler
from m4.models.vopt.configuration_vopt import VOPTConfig
from m4.training.utils import (
    compute_perceiver_tflops_per_batch_per_gpu,
    compute_tflops_per_batch_per_gpu,
    deepspeed_gathered_parameters_context_manager,
    freeze_model,
)
from m4.utils import logging


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "facebook/opt-350m"
_CONFIG_FOR_DOC = "VOPTConfig"
_TOKENIZER_FOR_DOC = "GPT2Tokenizer"

# Base model docstring
_EXPECTED_OUTPUT_SHAPE = [1, 8, 1024]

# SequenceClassification docstring
_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION = "ArthurZ/opt-350m-dummy-sc"
_SEQ_CLASS_EXPECTED_LOSS = 1.71
_SEQ_CLASS_EXPECTED_OUTPUT = "'LABEL_0'"


OPT_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "facebook/opt-125m",
    "facebook/opt-350m",
    "facebook/opt-1.3b",
    "facebook/opt-2.7b",
    "facebook/opt-6.7b",
    "facebook/opt-13b",
    "facebook/opt-30b",
    # See all OPT models at https://huggingface.co/models?filter=opt
]


class SwiGLUActivation(nn.Module):
    def __init__(self, in_features: int, out_features: int):
        super().__init__()
        self.gate = nn.Linear(in_features, out_features, bias=False)

    def forward(self, hidden_states_to_gate, hidden_states):
        gate = self.gate(hidden_states)
        return nn.functional.silu(gate) * hidden_states_to_gate


# Taken from LLaMA codebase
class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float()).type_as(x)
        return output * self.weight


def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
    """
    Make causal mask used for bi-directional self-attention.
    """
    bsz, tgt_len = input_ids_shape
    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min))
    mask_cond = torch.arange(mask.size(-1))
    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
    mask = mask.to(dtype)

    if past_key_values_length > 0:
        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)


def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
    """
    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
    """
    bsz, src_len = mask.size()
    tgt_len = tgt_len if tgt_len is not None else src_len

    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)

    inverted_mask = 1.0 - expanded_mask

    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)


class OPTLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int):
        # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
        # and adjust num_embeddings appropriately. Other models don't have this hack
        self.offset = 2
        super().__init__(num_embeddings + self.offset, embedding_dim)

    def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int = 0):
        """`input_ids_shape` is expected to be [bsz x seqlen]."""
        attention_mask = attention_mask.long()

        # create positions depending on attention_mask
        positions = (torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask).long() - 1

        # cut positions if `past_key_values_length` is > 0
        positions = positions[:, past_key_values_length:]

        return super().forward(positions + self.offset)


class OPTAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_cross_attention=False,
        config=None,
        qk_layer_norms=False,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder

        self.is_cross_attention = is_cross_attention

        if self.is_cross_attention:
            kv_input_dim = self.hidden_size if not hasattr(config, "vision_embed_dim") else config.vision_embed_dim
            self.k_proj = nn.Linear(kv_input_dim, embed_dim, bias=bias)
            self.v_proj = nn.Linear(kv_input_dim, embed_dim, bias=bias)

            self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        else:
            self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
            self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
            self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

        self.qk_layer_norms = qk_layer_norms
        if self.qk_layer_norms and config.rms_norm:
            self.q_layer_norm = RMSNorm(self.head_dim, eps=1e-6)
            self.k_layer_norm = RMSNorm(self.head_dim, eps=1e-6)
        elif self.qk_layer_norms:
            self.q_layer_norm = nn.LayerNorm(self.head_dim)
            self.k_layer_norm = nn.LayerNorm(self.head_dim)

    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = self.is_cross_attention or key_value_states is not None

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self._shape(self.q_proj(hidden_states), -1, bsz)
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        if self.qk_layer_norms:
            query_states = self.q_layer_norm(query_states)
            key_states = self.k_layer_norm(key_states)

        src_len = key_states.size(2)

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
        if layer_head_mask is not None:
            if layer_head_mask.size() != (self.num_heads,):
                raise ValueError(
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.size()}"
                )
            attention_mask = attention_mask.expand(-1, self.num_heads, -1, -1)
            attention_mask = attention_mask + layer_head_mask.view(1, -1, 1, 1)

        attn_output = nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.dropout,
        )

        attn_weights_reshaped = None
        logger.warning_once(
            "attn_weights are not extracted in scaled_dot_product_attention. The model returns None instead"
        )
        attn_output = attn_output.transpose(1, 2)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned aross GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value


class OPTDecoderLayer(nn.Module):
    def __init__(self, config: VOPTConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = OPTAttention(
            embed_dim=self.embed_dim,
            num_heads=config.num_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            config=config,
        )
        self.do_layer_norm_before = config.do_layer_norm_before
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.ffn_dim)
        self.fc2 = nn.Linear(config.ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            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`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
        """

        residual = hidden_states

        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
        if self.do_layer_norm_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        # 350m applies layer norm AFTER attention
        if not self.do_layer_norm_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Fully Connected
        hidden_states_shape = hidden_states.shape
        hidden_states = hidden_states.reshape(-1, hidden_states.size(-1))
        residual = hidden_states

        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
        if self.do_layer_norm_before:
            hidden_states = self.final_layer_norm(hidden_states)

        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)

        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        hidden_states = (residual + hidden_states).view(hidden_states_shape)

        # 350m applies layer norm AFTER attention
        if not self.do_layer_norm_before:
            hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


class VOPTGatedAttentionLayer(nn.Module):
    def __init__(self, config: VOPTConfig):
        """
        Note: Based on `tr_101_cm401xPMD09_nobias`, setting the biases to False in all of the nn.Linear for the gated cross attention.
        Provide a small stability gain at opt-13b scale.
        """
        super().__init__()
        self.embed_dim = config.hidden_size
        self.cross_attn = OPTAttention(
            embed_dim=self.embed_dim,
            num_heads=config.num_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            config=config,
            is_cross_attention=True,
            bias=False,
            qk_layer_norms=config.qk_layer_norms,
        )
        self.do_layer_norm_before = config.do_layer_norm_before
        self.normformer_layer_norms = config.normformer_layer_norms
        self.dropout = config.dropout
        if config.cross_layer_activation_function == "swiglu":
            # We cannot put `SwiGLUActivation` in `ACT2FN` because it takes two arguments (`in_features` and
            # `out_features`) that we don't know until entering this module.
            self.activation_fn = SwiGLUActivation(self.embed_dim, config.ffn_dim)
        else:
            self.activation_fn = ACT2FN[config.cross_layer_activation_function]

        if config.rms_norm:
            self.self_attn_layer_norm = RMSNorm(self.embed_dim, eps=1e-6)
        else:
            self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)

        if self.normformer_layer_norms:
            self.self_attn_post_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.ffn_dim, bias=False)
        self.fc2 = nn.Linear(config.ffn_dim, self.embed_dim, bias=False)

        if config.rms_norm:
            self.final_layer_norm = RMSNorm(self.embed_dim, eps=1e-6)
        else:
            self.final_layer_norm = nn.LayerNorm(self.embed_dim)

        if self.normformer_layer_norms:
            self.mlp_post_layer_norm = nn.LayerNorm(config.ffn_dim)

        self.act_cross_attn = nn.Tanh()
        self.act_dense = nn.Tanh()

        if config.alpha_initializer == "zeros":
            if config.alpha_type == "vector":
                self.alpha_cross_attn = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
                self.alpha_dense = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
            elif config.alpha_type == "float":
                self.alpha_cross_attn = nn.Parameter(torch.zeros(1))
                self.alpha_dense = nn.Parameter(torch.zeros(1))
            else:
                raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")

        elif config.alpha_initializer == "ones":
            if config.alpha_type == "vector":
                self.alpha_cross_attn = nn.Parameter(torch.ones(1, 1, self.embed_dim))
                self.alpha_dense = nn.Parameter(torch.ones(1, 1, self.embed_dim))
            elif config.alpha_type == "float":
                self.alpha_cross_attn = nn.Parameter(torch.ones(1))
                self.alpha_dense = nn.Parameter(torch.ones(1))
            else:
                raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")

        elif config.alpha_initializer in {"normal", "gaussian", "random"}:
            if config.alpha_type == "vector":
                self.alpha_cross_attn = nn.Parameter(
                    torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, self.embed_dim))
                )
                self.alpha_dense = nn.Parameter(
                    torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, self.embed_dim))
                )
            elif config.alpha_type == "float":
                self.alpha_cross_attn = nn.Parameter(
                    torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1))
                )
                self.alpha_dense = nn.Parameter(torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1)))
            else:
                raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")

        else:
            raise NotImplementedError(f"Alpha initialization scheme {config.alpha_initializer} not yet implemented!")

        assert hasattr(self, "alpha_cross_attn") and hasattr(self, "alpha_dense")

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        image_hidden_states: Optional[torch.Tensor] = None,
        image_attention_mask: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            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`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
        """
        if image_hidden_states is None:
            raise ValueError(
                "`image_hidden_states` is required for VOPT cross attention module which are visual features to be"
                " conditioned on."
            )

        if past_key_value is not None:
            raise NotImplementedError("Past key value states are not implemented for VOPT cross attention module.")

        residual = hidden_states

        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
        if self.do_layer_norm_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.cross_attn(
            hidden_states=hidden_states,
            key_value_states=image_hidden_states,
            attention_mask=image_attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        if self.normformer_layer_norms:
            hidden_states = self.self_attn_post_layer_norm(hidden_states)
        hidden_states = residual + self.act_cross_attn(self.alpha_cross_attn) * hidden_states

        # 350m applies layer norm AFTER attention
        if not self.do_layer_norm_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Fully Connected
        hidden_states_shape = hidden_states.shape
        hidden_states = hidden_states.reshape(-1, hidden_states.size(-1))
        residual = hidden_states

        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
        if self.do_layer_norm_before:
            hidden_states = self.final_layer_norm(hidden_states)

        hidden_states_to_gate = self.fc1(hidden_states)
        if isinstance(self.activation_fn, SwiGLUActivation):
            hidden_states = self.activation_fn(hidden_states_to_gate, hidden_states)
        else:
            hidden_states = self.activation_fn(hidden_states_to_gate)

        if self.normformer_layer_norms:
            hidden_states = self.mlp_post_layer_norm(hidden_states)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        hidden_states = (residual + self.act_dense(self.alpha_dense) * hidden_states).view(hidden_states_shape)

        # 350m applies layer norm AFTER attention
        if not self.do_layer_norm_before:
            hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


OPT_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 ([`VOPTConfig`]):
            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 OPT Model outputting raw hidden-states without any specific head on top.",
    OPT_START_DOCSTRING,
)
class VOPTPreTrainedModel(VLOOMPreTrainedModelBase):
    config_class = VOPTConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["OPTDecoderLayer", "VOPTGatedAttentionLayer", "CLIPEncoderLayer"]
    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]

    def _init_weights(self, module):
        def init_a_linear(module, mean=0.0, std=self.config.init_std):
            with ContextManagers(deepspeed_gathered_parameters_context_manager(module.weight, modify=True)):
                module.weight.data.normal_(mean=mean, std=std)
                if module.bias is not None:
                    with ContextManagers(deepspeed_gathered_parameters_context_manager(module.bias, modify=True)):
                        module.bias.data.zero_()

        if isinstance(module, VOPTGatedAttentionLayer):
            for sub_module_name, sub_module in module.named_modules():
                if isinstance(sub_module, nn.Linear):
                    if "fc2" in sub_module_name:
                        factor = 2 * self.config.num_hidden_layers
                    else:
                        factor = 1.0
                    init_a_linear(sub_module, std=(0.4 / (sub_module.in_features * factor)) ** 0.5)
        elif isinstance(module, PerceiverResampler):
            with ContextManagers(deepspeed_gathered_parameters_context_manager(module.latents, modify=True)):
                module.latents.data.normal_(mean=0.0, std=(1.0 / self.config.vision_embed_dim) ** 0.5)
            for sub_module_name, sub_module in module.named_modules():
                if isinstance(sub_module, nn.Linear):
                    if "c_proj" in sub_module_name:
                        factor = 2 * self.config.num_hidden_layers
                    else:
                        factor = 1.0
                    init_a_linear(sub_module, std=(0.4 / (self.config.vision_embed_dim * factor)) ** 0.5)
        elif isinstance(module, nn.Embedding):
            with ContextManagers(deepspeed_gathered_parameters_context_manager(module.weight, modify=True)):
                module.weight.data.normal_(mean=0.0, std=(1.0 / self.config.hidden_size) ** 0.5)
                if module.padding_idx is not None:
                    module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, DecoupledLinear):
            if hasattr(module, "additional_fc"):
                init_a_linear(module.additional_fc, std=(1.0 / (module.additional_fc.in_features)) ** 0.5)

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (VOPTDecoder)):
            module.gradient_checkpointing = value

    @classmethod
    def override_vision_model_wrapper(cls, model, config, vision_model_name, vision_model_params, torch_dtype):
        # this can be called via from_pretrained from a class w/ head or w/o head so we extract the beheaded model version
        beheaded_model = model.model if hasattr(model, "model") else model
        cls.override_vision_model(beheaded_model.decoder, vision_model_name, vision_model_params, torch_dtype)
        beheaded_model.freeze_relevant_params(config)


OPT_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 [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        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 [`OPTTokenizer`]. 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.
        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        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.
"""


class VOPTDecoder(VOPTPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`OPTDecoderLayer`]

    Args:
        config: VOPTConfig
    """

    def __init__(self, config: VOPTConfig, vision_model=None):
        super().__init__(config)
        self.config = config
        self.dropout = config.dropout
        self.layerdrop = config.layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_position_embeddings
        self.vocab_size = config.vocab_size

        self.embed_tokens = DecoupledEmbedding(
            num_embeddings=config.vocab_size,
            num_additional_embeddings=config.additional_vocab_size,
            embedding_dim=config.word_embed_proj_dim,
            partially_freeze=config.freeze_text_layers,
            padding_idx=self.padding_idx,
        )
        self.embed_positions = OPTLearnedPositionalEmbedding(config.max_position_embeddings, config.hidden_size)

        # Load an uninitialized model and later in from_pretrained will load the pre-trained model -
        # this solves the losing of weights in `from_pretrained` on the main model
        self.vision_model = vision_model

        # Perceiver Resampler
        if config.use_resampler:
            self.perceiver_resampler = PerceiverResampler(
                self.config,
                self.config.vision_embed_dim,
                config.resampler_depth,
                config.resampler_n_heads,
                config.resampler_head_dim,
                config.resampler_n_latents,
            )

        if config.word_embed_proj_dim != config.hidden_size:
            self.project_in = nn.Linear(config.word_embed_proj_dim, config.hidden_size, bias=False)
        else:
            self.project_in = None

        self.layers = nn.ModuleList([OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)])

        self.cross_layer_interval = config.cross_layer_interval
        num_cross_layers = config.num_hidden_layers // self.cross_layer_interval
        self.gated_cross_attn_layers = nn.ModuleList(
            [VOPTGatedAttentionLayer(config) for i in range(num_cross_layers)]
        )
        self.gradient_checkpointing = False

        # Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility
        # with checkpoints that have been fine-tuned before transformers v4.20.1
        # see https://github.com/facebookresearch/metaseq/pull/164
        if config.do_layer_norm_before and not config._remove_final_layer_norm:
            self.final_layer_norm = nn.LayerNorm(config.hidden_size)
        else:
            self.final_layer_norm = None

        if config.word_embed_proj_dim != config.hidden_size:
            self.project_out = nn.Linear(config.hidden_size, config.word_embed_proj_dim, bias=False)
        else:
            self.project_out = None

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
        # create causal mask
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = None
        if input_shape[-1] > 1:
            combined_attention_mask = _make_causal_mask(
                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
            ).to(inputs_embeds.device)

        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
                inputs_embeds.device
            )
            combined_attention_mask = (
                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
            )

        return combined_attention_mask

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_embeddings: Optional[torch.FloatTensor] = None,
        image_attention_mask: Optional[torch.Tensor] = None,
        crossblock_head_mask: Optional[torch.Tensor] = None,  # TOFO (ls): check if this is needed
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        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 [`OPTTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            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)
            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            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

                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.
            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.
        """
        device = input_ids.device if input_ids is not None else inputs_embeds.device

        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0

        if pixel_values is not None and image_embeddings is not None:
            raise ValueError("You cannot specify both pixel_values and image_embeddings at the same time")
        elif pixel_values is not None:
            pixel_values = pixel_values.to(dtype=self.dtype, device=input_ids.device)  # fp16 compatibility
            batch_size, num_images = pixel_values.size(0), pixel_values.size(1)
            pixel_values = pixel_values.contiguous().view(batch_size * num_images, *pixel_values.shape[2:])
            # Get sequence from the vision encoder
            image_hidden_states = self.vision_model(pixel_values=pixel_values).last_hidden_state
        elif image_embeddings is not None:
            batch_size, num_images, image_seq_len, image_hidden_size = image_embeddings.size()
            image_hidden_states = image_embeddings.to(dtype=self.dtype, device=input_ids.device)
            image_hidden_states = image_hidden_states.view(batch_size * num_images, image_seq_len, image_hidden_size)

        if self.config.use_resampler:
            image_hidden_states = self.perceiver_resampler(image_hidden_states)
        image_seq_len, image_hidden_size = image_hidden_states.size(1), image_hidden_states.size(2)
        image_hidden_states = image_hidden_states.view(batch_size, num_images * image_seq_len, image_hidden_size)
        # Make image_attention_mask compatible with hidden states
        text_seq_len = image_attention_mask.size(1)
        image_attention_mask = image_attention_mask.unsqueeze(-1)
        image_attention_mask = image_attention_mask.repeat(1, 1, 1, image_seq_len)
        image_attention_mask = image_attention_mask.view(batch_size, text_seq_len, num_images * image_seq_len)

        if image_hidden_states is not None:
            image_batch_size, image_sequence_length, _ = image_hidden_states.size()
            image_hidden_shape = (image_batch_size, image_sequence_length)
            if image_attention_mask is None:
                image_attention_mask = torch.ones(image_hidden_shape, device=device)
            image_attention_mask = self.invert_attention_mask(image_attention_mask)
        else:
            image_attention_mask = None

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

        # embed positions
        if attention_mask is None:
            attention_mask = torch.ones(inputs_embeds.shape[:2], dtype=torch.bool, device=inputs_embeds.device)
        pos_embeds = self.embed_positions(attention_mask, past_key_values_length)

        attention_mask = self._prepare_decoder_attention_mask(
            attention_mask, input_shape, inputs_embeds, past_key_values_length
        )

        if self.project_in is not None:
            inputs_embeds = self.project_in(inputs_embeds)

        hidden_states = inputs_embeds + pos_embeds

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        next_decoder_cache = () if use_cache else None

        # check if head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip([head_mask], ["head_mask"]):
            if attn_mask is not None:
                if attn_mask.size()[0] != (len(self.layers)):
                    raise ValueError(
                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                        f" {head_mask.size()[0]}."
                    )

        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None
            layer_head_mask = head_mask[idx] if head_mask is not None else None

            def vblock(
                main_block,
                hidden_states,
                attention_mask,
                layer_head_mask,
                past_key_value,
                image_hidden_states,
                image_attention_mask,
                output_attentions,
                use_cache,
                layer_idx,
                cross_layer_interval,
                gated_cross_attn_layers,
            ):
                # TODO(ls): Add cross attention values to respective lists
                if layer_idx % cross_layer_interval == 0:
                    xblock = gated_cross_attn_layers[layer_idx // cross_layer_interval]
                    outputs = xblock(
                        hidden_states,
                        attention_mask=attention_mask,
                        layer_head_mask=layer_head_mask,
                        image_hidden_states=image_hidden_states,
                        image_attention_mask=image_attention_mask,
                        output_attentions=output_attentions,
                        use_cache=use_cache,
                        past_key_value=None,  # not implemented
                    )
                    hidden_states = outputs[0]

                layer_outputs = main_block(
                    hidden_states,
                    attention_mask=attention_mask,
                    layer_head_mask=layer_head_mask,
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

                return layer_outputs

            if self.gradient_checkpointing and self.training:
                past_key_value = None
                if use_cache:
                    logger.warning_once(
                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                    )
                    use_cache = False

                layer_outputs = torch.utils.checkpoint.checkpoint(
                    vblock,
                    decoder_layer,
                    hidden_states,
                    attention_mask,
                    layer_head_mask,
                    past_key_value,
                    image_hidden_states,
                    image_attention_mask,
                    output_attentions,
                    use_cache,
                    idx,
                    self.cross_layer_interval,
                    self.gated_cross_attn_layers,
                )
            else:
                layer_outputs = vblock(
                    decoder_layer,
                    hidden_states,
                    attention_mask=attention_mask,
                    layer_head_mask=layer_head_mask,
                    past_key_value=past_key_value,
                    image_hidden_states=image_hidden_states,
                    image_attention_mask=image_attention_mask,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    layer_idx=idx,
                    cross_layer_interval=self.cross_layer_interval,
                    gated_cross_attn_layers=self.gated_cross_attn_layers,
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        if self.final_layer_norm is not None:
            hidden_states = self.final_layer_norm(hidden_states)

        if self.project_out is not None:
            hidden_states = self.project_out(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


@add_start_docstrings(
    "The bare OPT Model outputting raw hidden-states without any specific head on top.",
    OPT_START_DOCSTRING,
)
class VOPTModel(VOPTPreTrainedModel):
    def __init__(self, config: VOPTConfig, vision_model=None):
        super().__init__(config)
        self.decoder = VOPTDecoder(config, vision_model=vision_model)

        # Initialize weights and apply final processing
        self.post_init()

        self.freeze_relevant_params(config)

    def freeze_relevant_params(self, config=None):
        if config is None:
            config = self.config

        if config.freeze_text_layers:
            self.freeze_text_layers(config.freeze_text_module_exceptions)

        if config.freeze_vision_layers:
            freeze_model(self.decoder.vision_model, module_exceptions=config.freeze_vision_module_exceptions)

    def freeze_text_layers(self, module_exceptions):
        for module in [self.decoder.embed_positions, self.decoder.layers]:
            freeze_model(module, module_exceptions=module_exceptions)

        if self.decoder.project_out is not None:
            freeze_model(self.decoder.project_out, module_exceptions=module_exceptions)

        if self.decoder.final_layer_norm is not None:
            freeze_model(self.decoder.final_layer_norm, module_exceptions=module_exceptions)

    def get_input_embeddings(self):
        return self.decoder.embed_tokens

    def set_input_embeddings(self, value):
        self.decoder.embed_tokens = value

    def get_decoder(self):
        return self.decoder

    @add_start_docstrings_to_model_forward(OPT_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        processor_class=_TOKENIZER_FOR_DOC,
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=BaseModelOutputWithPast,
        config_class=_CONFIG_FOR_DOC,
        expected_output=_EXPECTED_OUTPUT_SHAPE,
    )
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_embeddings: Optional[torch.FloatTensor] = None,
        image_attention_mask: Optional[torch.Tensor] = None,
        crossblock_head_mask: Optional[torch.Tensor] = None,  # TOFO (ls): check if this is needed
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            pixel_values=pixel_values,
            image_embeddings=image_embeddings,
            image_attention_mask=image_attention_mask,
            crossblock_head_mask=crossblock_head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        if not return_dict:
            return decoder_outputs

        return BaseModelOutputWithPast(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            hidden_states=decoder_outputs.hidden_states,
            attentions=decoder_outputs.attentions,
        )


class VOPTForCausalLM(VOPTPreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]

    def __init__(self, config, vision_model=None):
        super().__init__(config)

        # Initialize LM head first so that it is not directly offloaded to the CPU/disk
        # the lm_head weight is automatically tied to the embed tokens weight
        self.lm_head = DecoupledLinear(
            in_features=config.word_embed_proj_dim,
            out_features=config.vocab_size,
            out_additional_features=config.additional_vocab_size,
            bias=False,
            partially_freeze=config.freeze_lm_head,
        )

        self.model = VOPTModel(config, vision_model=vision_model)
        # Initialize weights and apply final processing
        self.post_init()

    def tie_weights(self):
        """
        Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of DecoupledLinear and DecoupledEmbedding.
        """
        output_embeddings = self.get_output_embeddings()
        input_embeddings = self.get_input_embeddings()

        if getattr(self.config, "tie_word_embeddings", True):
            output_embeddings.weight = input_embeddings.weight
            if input_embeddings.num_additional_embeddings > 0:
                assert output_embeddings.out_additional_features == input_embeddings.num_additional_embeddings
                output_embeddings.additional_fc.weight = input_embeddings.additional_embedding.weight

        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
            output_embeddings.out_features = input_embeddings.num_embeddings
            if hasattr(output_embeddings, "out_additional_features") and hasattr(
                input_embeddings, "num_additional_embeddings"
            ):
                output_embeddings.out_additional_features = input_embeddings.num_additional_embeddings

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

    def set_input_embeddings(self, value):
        self.model.decoder.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model.decoder = decoder

    def get_decoder(self):
        return self.model.decoder

    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_embeddings: Optional[torch.FloatTensor] = None,
        image_attention_mask: Optional[torch.Tensor] = None,
        crossblock_head_mask: Optional[torch.Tensor] = 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, CausalLMOutputWithPast]:
        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 [`OPTTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            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)
            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            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)`. The two additional
                tensors are only required when the model is used as a decoder in a Sequence to Sequence model.

                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.
            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]`.
            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.

        Returns:

        Example:

        ```python
        >>> from transformers import GPT2Tokenizer, OPTForCausalLM

        >>> model = OPTForCausalLM.from_pretrained("facebook/opt-350m")
        >>> tokenizer = GPT2Tokenizer.from_pretrained("facebook/opt-350m")

        >>> prompt = "Hey, are you consciours? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
        ```"""

        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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            pixel_values=pixel_values,
            image_embeddings=image_embeddings,
            image_attention_mask=image_attention_mask,
            crossblock_head_mask=crossblock_head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        logits = self.lm_head(outputs[0]).contiguous()

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            if attention_mask is not None:
                shift_attention_mask = attention_mask[..., 1:]
                shift_logits = logits[..., :-1, :][shift_attention_mask != 0].contiguous()
                shift_labels = labels[..., 1:][shift_attention_mask != 0].contiguous()
            else:
                shift_logits = logits[..., :-1, :].contiguous()
                shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            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=None, **kwargs):
        inputs = prepare_inputs_for_generation(input_ids, past=past, **kwargs)
        unwanted_kwargs = ["position_ids", "token_type_ids"]
        for kwarg in unwanted_kwargs:
            inputs.pop(kwarg, None)
        return inputs

    @staticmethod
    def _expand_inputs_for_generation(
        *args,
        **model_kwargs,
    ):
        return expand_inputs_for_generation(*args, **model_kwargs)

    @staticmethod
    def _update_model_kwargs_for_generation(outputs, model_kwargs, is_encoder_decoder=False):
        return update_model_kwargs_for_generation(outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder)

    @staticmethod
    def _reorder_cache(past, beam_idx):
        reordered_past = ()
        for layer_past in past:
            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
        return reordered_past

    def get_model_tflops_per_batch_per_gpu(self, hparams, data_param, tokenizer, max_num_images):
        config_vl_model = self.config

        language_embed_size = config_vl_model.hidden_size
        num_language_layers = config_vl_model.num_hidden_layers
        ffn_inner_size = config_vl_model.ffn_dim

        vision_config = self.model.decoder.vision_model.config
        if hasattr(vision_config, "vision_config"):
            vision_config = vision_config.vision_config

        # Get vision model blocks infos
        vision_patch_size = vision_config.patch_size
        vision_hidden_size = vision_config.hidden_size
        num_vision_layers = vision_config.num_hidden_layers
        # The +1 is for the CLS token
        single_image_seq_len = (vision_config.image_size // vision_patch_size) ** 2 + 1
        vision_exp_factor = vision_config.intermediate_size // vision_hidden_size

        # Get language and cross-att blocks infos
        num_cross_attn_layers = num_language_layers // config_vl_model.cross_layer_interval
        language_seq_len = data_param.max_seq_len
        language_exp_factor = (ffn_inner_size // language_embed_size) if ffn_inner_size is not None else 4
        cross_att_exp_factor = (ffn_inner_size // language_embed_size) if ffn_inner_size is not None else 4
        k_v_cross_attn_seq_len = (
            (self.config.resampler_n_latents * max_num_images)
            if self.config.use_resampler
            else (single_image_seq_len * max_num_images)
        )

        language_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
            num_layers=num_language_layers,
            batch_size=hparams.batch_size_per_gpu,
            q_seq_len=language_seq_len,
            k_seq_len=language_seq_len,
            hidden_size=language_embed_size,
            kv_in_dim=language_embed_size,
            ff_exp_factor=language_exp_factor,
            grad_acc_size=hparams.grad_acc_size,
            swiglu=False,
            vocab_size=tokenizer.vocab_size,
            count_backward=True,  # Always True regardless of freezing, because gradients are computed for cross-attentions
            use_grad_checkpointing=hparams.gradient_checkpointing,
        )
        cross_attention_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
            num_layers=num_cross_attn_layers,
            batch_size=hparams.batch_size_per_gpu,
            q_seq_len=language_seq_len,
            k_seq_len=k_v_cross_attn_seq_len,
            hidden_size=language_embed_size,
            kv_in_dim=vision_hidden_size,
            ff_exp_factor=cross_att_exp_factor,
            grad_acc_size=hparams.grad_acc_size,
            swiglu=self.config.cross_layer_activation_function == "swiglu",
            vocab_size=None,
            count_backward=True,
            use_grad_checkpointing=hparams.gradient_checkpointing,
        )
        vision_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
            num_layers=num_vision_layers,
            batch_size=hparams.batch_size_per_gpu * max_num_images,
            q_seq_len=single_image_seq_len,
            k_seq_len=single_image_seq_len,
            hidden_size=vision_hidden_size,
            kv_in_dim=vision_hidden_size,
            ff_exp_factor=vision_exp_factor,
            grad_acc_size=hparams.grad_acc_size,
            swiglu=False,
            vocab_size=None,
            count_backward=not hparams.model_params["freeze_vision_layers"],
            use_grad_checkpointing=hparams.gradient_checkpointing,
        )
        if self.config.use_resampler:
            perceiver_tflops_per_batch_per_gpu = compute_perceiver_tflops_per_batch_per_gpu(
                num_layers=self.config.resampler_depth,
                batch_size=hparams.batch_size_per_gpu * max_num_images,
                q_seq_len=self.config.resampler_n_latents,
                vision_embed_seq_len=single_image_seq_len,
                q_k_v_input_dim=vision_hidden_size,
                attention_hidden_size=self.config.resampler_n_heads * self.config.resampler_head_dim,
                ff_exp_factor=cross_att_exp_factor,
                count_backward=True,
                use_grad_checkpointing=hparams.gradient_checkpointing,
            )
            flop_count = (
                language_tflops_per_batch_per_gpu
                + cross_attention_tflops_per_batch_per_gpu
                + vision_tflops_per_batch_per_gpu
                + perceiver_tflops_per_batch_per_gpu
            )
        else:
            flop_count = (
                language_tflops_per_batch_per_gpu
                + cross_attention_tflops_per_batch_per_gpu
                + vision_tflops_per_batch_per_gpu
            )
        return flop_count