m4/models/vbloom/modeling_vbloom.py

# coding=utf-8
# Copyright 2022 HuggingFace Inc. team and BigScience workshop.
#
# 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 BLOOM model."""

import math
import warnings
from typing import Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss, LayerNorm
from torch.nn import functional as F
from transformers.file_utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
)
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions

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.vbloom.configuration_vbloom import VBloomConfig
from m4.training.utils import (
    compute_perceiver_tflops_per_batch_per_gpu,
    compute_tflops_per_batch_per_gpu,
    freeze_model,
)
from m4.utils import logging


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "bigscience/bloom-560m"
_CONFIG_FOR_DOC = "VBloomConfig"
_TOKENIZER_FOR_DOC = "BloomTokenizerFast"

BLOOM_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "bigscience/bigscience-small-testing",
    "bigscience/bloom-560m",
    "bigscience/bloom-1b1",
    "bigscience/bloom-1b7",
    "bigscience/bloom-3b",
    "bigscience/bloom-7b1",
    "bigscience/bloom",
]


def _make_causal_mask(
    input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
) -> torch.BoolTensor:
    """
    Make causal mask used for self-attention.
    """
    batch_size, target_length = input_ids_shape
    mask = torch.empty((target_length, target_length + past_key_values_length), dtype=torch.bool, device=device)
    # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
    seq_ids = torch.arange(target_length, device=device)
    mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]

    if past_key_values_length > 0:
        mask[:, :past_key_values_length] = False

    expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
    return expanded_mask


def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
    """
    Expands attention_mask from `[batch_size, src_length]` to `[batch_size, 1, tgt_length, src_length]`.
    """
    batch_size, src_length = mask.shape
    tgt_length = tgt_length if tgt_length is not None else src_length

    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
    return expanded_mask.expand(batch_size, 1, tgt_length, src_length)


def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
    """
    Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
    relies on a translation invariance of softmax for quick implementation: with l being a tensor, and a fixed value
    `softmax(l+a) = softmax(l)`. Based on
    https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
    TODO @thomasw21 this doesn't work as nicely due to the masking strategy, and so masking varies slightly.

    Args:
    Returns tensor shaped (batch_size * num_heads, 1, max_seq_len)
        attention_mask (`torch.Tensor`):
            Token-wise attention mask, this should be of shape (batch_size, max_seq_len).
        num_heads (`int`, *required*):
            number of heads
        dtype (`torch.dtype`, *optional*, default=`torch.bfloat16`):
            dtype of the output tensor
    """
    batch_size, seq_length = attention_mask.shape
    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
    base = torch.tensor(
        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
    )
    powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
    slopes = torch.pow(base, powers)

    if closest_power_of_2 != num_heads:
        extra_base = torch.tensor(
            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
        )
        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
        extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32)
        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)

    # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
    # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
    # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
    # => the query_length dimension will then be broadcasted correctly
    # This is more or less identical to T5's relative position bias:
    # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
    arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
    alibi = slopes[..., None] * arange_tensor
    return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)


def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor:
    """
    Dropout add function

    Args:
        x (`torch.tensor`, *required*):
            input tensor
        residual (`torch.tensor`, *required*):
            esidual tensor
        prob (`float`, *required*):
            dropout probability
        training (`bool`, *required*):
            training mode
    """
    out = F.dropout(x, p=prob, training=training)
    out = residual + out
    return out


def bloom_gelu_forward(x: torch.Tensor) -> torch.Tensor:
    """
    Custom bias GELU function. Adapted from Megatron-DeepSpeed code. Here we use a simple implementation (inference) to
    make the model jitable.

    Args:
        x (`torch.tensor`, *required*):
            input hidden states
    """
    return x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))


def bloom_gelu_back(g: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
    """
    gradient of tanh approximation of gelu gradient of actual gelu is: 0.5 * (1. + torch.erf(x * 0.70710678)) +
    0.3989423 * x * torch.exp(-0.5 * x * x)

    Args:
        g (`torch.tensor`, *required*):
            gradient output tensor
        x (`torch.tensor`, *required*):
            input tensor
    """
    x = x[0]  # x is a tuple of 1 element, needs to unpack it first
    tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
    # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
    ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
    return ff * g


class GeLUFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input: torch.Tensor) -> torch.Tensor:
        ctx.save_for_backward(input)
        return bloom_gelu_forward(input)

    @staticmethod
    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
        input = ctx.saved_tensors
        tmp = bloom_gelu_back(grad_output, input)
        return tmp


class BloomGelu(nn.Module):
    """
    BloomBiasGelu wrapper function that make use of the simple function on inference mode to make the model
    torchscriptable and use the autograd function in training mode to get the accurate results of the gradients Partly
    copied from Megatron-DeepSpeed code and adapted for our needs

    See here why autograd functions are not torchscriptable: https://github.com/pytorch/pytorch/issues/22329
    """

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

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        if self.training:
            return GeLUFunction.apply(x)
        else:
            return bloom_gelu_forward(x)


class BloomAttention(nn.Module):
    def __init__(self, config: VBloomConfig, is_cross_attention=False):
        super().__init__()

        self.pretraining_tp = config.pretraining_tp
        self.slow_but_exact = config.slow_but_exact

        self.hidden_size = config.hidden_size
        self.num_heads = config.n_head
        self.head_dim = self.hidden_size // self.num_heads
        self.split_size = self.hidden_size
        self.hidden_dropout = config.hidden_dropout

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

        # Layer-wise attention scaling
        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
        self.beta = 1.0

        self.is_cross_attention = is_cross_attention

        if self.is_cross_attention:
            self.query = nn.Linear(self.hidden_size, 1 * self.hidden_size, bias=True)
            kv_input_dim = self.hidden_size if not hasattr(config, "vision_embed_dim") else config.vision_embed_dim
            self.key_value = nn.Linear(kv_input_dim, 2 * self.hidden_size, bias=True)
        else:
            self.query_key_value = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=True)

        self.dense = nn.Linear(self.hidden_size, self.hidden_size)
        self.attention_dropout = nn.Dropout(config.attention_dropout)

        if self.is_cross_attention:
            # The alpha stuff
            self.act = nn.Tanh()

            if config.alpha_initializer == "zeros":
                if config.alpha_type == "vector":
                    self.alpha_cross_attn = nn.Parameter(torch.zeros(1, 1, self.hidden_size))
                elif config.alpha_type == "float":
                    self.alpha_cross_attn = 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.hidden_size))
                elif config.alpha_type == "float":
                    self.alpha_cross_attn = 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.hidden_size))
                    )
                elif config.alpha_type == "float":
                    self.alpha_cross_attn = 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!"
                )

    def _split_heads(self, fused_qkv: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Split the last dimension into (num_heads, head_dim) without making any copies, results share same memory
        storage as `fused_qkv`

        Args:
            fused_qkv (`torch.tensor`, *required*): [batch_size, seq_length, num_heads * 3 * head_dim]

        Returns:
            query: [batch_size, seq_length, num_heads, head_dim] key: [batch_size, seq_length, num_heads, head_dim]
            value: [batch_size, seq_length, num_heads, head_dim]
        """
        batch_size, seq_length, n_times_hidden_size = fused_qkv.shape
        n = int(n_times_hidden_size / self.hidden_size)
        fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads, n, self.head_dim)
        outputs = ()
        for i in range(n):
            outputs += (fused_qkv[..., i, :],)
        return outputs

    def _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
        """
        Merge heads together over the last dimenstion

        Args:
            x: (`torch.tensor`, *required*): [batch_size * num_heads, seq_length, head_dim]

        Returns:
            torch.tensor: [batch_size, seq_length, num_heads * head_dim]
        """
        # What we want to achieve is:
        # batch_size * num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads * head_dim
        batch_size_and_num_heads, seq_length, _ = x.shape
        batch_size = batch_size_and_num_heads // self.num_heads

        # First view to decompose the batch size
        # batch_size * num_heads, seq_length, head_dim -> batch_size, num_heads, seq_length, head_dim
        x = x.view(batch_size, self.num_heads, seq_length, self.head_dim)

        # batch_size, num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads, head_dim
        x = x.permute(0, 2, 1, 3)

        # batch_size, seq_length, num_heads, head_dim -> batch_size, seq_length, num_heads * head_dim
        return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: torch.Tensor,
        alibi: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        head_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: bool = False,
        output_attentions: bool = False,
    ):
        if not self.is_cross_attention:
            fused_qkv = self.query_key_value(hidden_states)  # [batch_size, seq_length, 3 x hidden_size]

            # 3 x [batch_size, seq_length, num_heads, head_dim]
            (query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
        else:
            if encoder_hidden_states is not None:
                attention_mask = encoder_attention_mask
            q = self.query(hidden_states)
            kv = self.key_value(encoder_hidden_states)

            query_layer = self._split_heads(q)[0]
            key_layer, value_layer = self._split_heads(kv)

        batch_size, q_length, _, _ = query_layer.shape
        _, kv_length, _, _ = key_layer.shape

        query_layer = query_layer.transpose(1, 2).reshape(batch_size * self.num_heads, q_length, self.head_dim)
        key_layer = key_layer.permute(0, 2, 3, 1).reshape(batch_size * self.num_heads, self.head_dim, kv_length)
        value_layer = value_layer.transpose(1, 2).reshape(batch_size * self.num_heads, kv_length, self.head_dim)
        if layer_past is not None:
            past_key, past_value = layer_past
            # concatenate along seq_length dimension:
            #  - key: [batch_size * self.num_heads, head_dim, kv_length]
            #  - value: [batch_size * self.num_heads, kv_length, head_dim]
            key_layer = torch.cat((past_key, key_layer), dim=2)
            value_layer = torch.cat((past_value, value_layer), dim=1)
            _, _, kv_length = key_layer.shape

        if use_cache is True:
            present = (key_layer, value_layer)
        else:
            present = None

        # [batch_size * num_heads, q_length, kv_length]
        # we use `torch.Tensor.baddbmm` instead of `torch.baddbmm` as the latter isn't supported by TorchScript v1.11
        if alibi is None:
            alibi = torch.empty(
                batch_size * self.num_heads, q_length, kv_length, dtype=query_layer.dtype, device=query_layer.device
            )

        matmul_result = alibi.baddbmm(
            batch1=query_layer,
            batch2=key_layer,
            beta=0.0 if self.is_cross_attention else self.beta,
            alpha=self.inv_norm_factor,
        )

        # change view to [batch_size, num_heads, q_length, kv_length]
        attention_scores = matmul_result.view(batch_size, self.num_heads, q_length, kv_length)

        # cast attention scores to fp32, compute scaled softmax and cast back to initial dtype - [batch_size, num_heads, q_length, kv_length]
        input_dtype = attention_scores.dtype
        # `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
        if input_dtype == torch.float16:
            attention_scores = attention_scores.to(torch.float)
        attn_weights = torch.masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
        attention_probs = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(input_dtype)

        # [batch_size, num_heads, q_length, kv_length]
        attention_probs = self.attention_dropout(attention_probs)

        if head_mask is not None:
            attention_probs = attention_probs * head_mask

        # change view [batch_size x num_heads, q_length, kv_length]
        attention_probs_reshaped = attention_probs.view(batch_size * self.num_heads, q_length, kv_length)

        # matmul: [batch_size * num_heads, q_length, head_dim]
        context_layer = torch.bmm(attention_probs_reshaped, value_layer)

        # change view [batch_size, num_heads, q_length, head_dim]
        context_layer = self._merge_heads(context_layer)

        # aggregate results across tp ranks. See here: https://github.com/pytorch/pytorch/issues/76232
        if self.pretraining_tp > 1 and self.slow_but_exact:
            slices = self.hidden_size / self.pretraining_tp
            output_tensor = torch.zeros_like(context_layer)
            for i in range(self.pretraining_tp):
                output_tensor = output_tensor + F.linear(
                    context_layer[:, :, int(i * slices) : int((i + 1) * slices)],
                    self.dense.weight[:, int(i * slices) : int((i + 1) * slices)],
                )
        else:
            output_tensor = self.dense(context_layer)

        if not self.is_cross_attention:
            output_tensor = dropout_add(output_tensor, residual, self.hidden_dropout, self.training)
        else:
            output_tensor = dropout_add(
                self.act(self.alpha_cross_attn) * output_tensor, residual, self.hidden_dropout, self.training
            )

        outputs = (output_tensor, present)
        if output_attentions:
            outputs += (attention_probs,)

        return outputs


class BloomMLP(nn.Module):
    def __init__(self, config: VBloomConfig, is_gated=False):
        super().__init__()
        hidden_size = config.hidden_size

        self.pretraining_tp = config.pretraining_tp
        self.slow_but_exact = config.slow_but_exact
        self.dense_h_to_4h = nn.Linear(hidden_size, 4 * hidden_size)
        self.gelu_impl = BloomGelu()
        self.dense_4h_to_h = nn.Linear(4 * hidden_size, hidden_size)
        self.hidden_dropout = config.hidden_dropout

        # The alpha stuff
        self.is_gated = is_gated
        if is_gated:
            self.act = nn.Tanh()

            if config.alpha_initializer == "zeros":
                if config.alpha_type == "vector":
                    self.alpha_dense = nn.Parameter(torch.zeros(1, 1, hidden_size))
                elif config.alpha_type == "float":
                    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_dense = nn.Parameter(torch.ones(1, 1, hidden_size))
                elif config.alpha_type == "float":
                    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_dense = nn.Parameter(
                        torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, hidden_size))
                    )
                elif config.alpha_type == "float":
                    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!"
                )

    def forward(self, hidden_states: torch.Tensor, residual: torch.Tensor) -> torch.Tensor:
        hidden_states = self.gelu_impl(self.dense_h_to_4h(hidden_states))

        if self.pretraining_tp > 1 and self.slow_but_exact:
            intermediate_output = torch.zeros_like(residual)
            slices = self.dense_4h_to_h.weight.shape[-1] / self.pretraining_tp
            for i in range(self.pretraining_tp):
                intermediate_output = intermediate_output + F.linear(
                    hidden_states[:, :, int(i * slices) : int((i + 1) * slices)],
                    self.dense_4h_to_h.weight[:, int(i * slices) : int((i + 1) * slices)],
                )
        else:
            intermediate_output = self.dense_4h_to_h(hidden_states)

        if not self.is_gated:
            output = dropout_add(intermediate_output, residual, self.hidden_dropout, self.training)
        else:
            output = dropout_add(
                self.act(self.alpha_dense) * intermediate_output, residual, self.hidden_dropout, self.training
            )

        return output


class BloomBlock(nn.Module):
    def __init__(self, config: VBloomConfig):
        super().__init__()
        hidden_size = config.hidden_size

        self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.num_heads = config.n_head
        self.self_attention = BloomAttention(config)
        self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

        self.mlp = BloomMLP(config)

        self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
        self.hidden_dropout = config.hidden_dropout

    def forward(
        self,
        hidden_states: torch.Tensor,
        alibi: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        head_mask: Optional[torch.Tensor] = None,
        use_cache: bool = False,
        output_attentions: bool = False,
    ):
        # hidden_states: [batch_size, seq_length, hidden_size]

        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.input_layernorm(hidden_states)

        # Layer norm post the self attention.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = hidden_states

        # Self attention.
        attn_outputs = self.self_attention(
            layernorm_output,
            residual,
            layer_past=layer_past,
            attention_mask=attention_mask,
            alibi=alibi,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )

        attention_output = attn_outputs[0]

        outputs = attn_outputs[1:]

        layernorm_output = self.post_attention_layernorm(attention_output)

        # Get residual
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = attention_output

        # MLP.
        output = self.mlp(layernorm_output, residual)

        if use_cache:
            outputs = (output,) + outputs
        else:
            outputs = (output,) + outputs[1:]

        return outputs  # hidden_states, present, attentions


class VBloomGatedCrossAttentionBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        hidden_size = config.hidden_size

        self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.num_heads = config.n_head
        self.cross_attention = BloomAttention(config, is_cross_attention=True)
        self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

        self.gated_mlp = BloomMLP(config, is_gated=True)

        self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
        self.hidden_dropout = config.hidden_dropout

    def forward(
        self,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        image_hidden_states: Optional[torch.Tensor] = None,
        image_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
        # hidden_states: [batch_size, seq_length, hidden_size]

        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.input_layernorm(hidden_states)

        # Layer norm post the self attention.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = hidden_states

        # Self attention.
        attn_outputs = self.cross_attention(
            layernorm_output,
            residual,
            alibi=None,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            encoder_hidden_states=image_hidden_states,
            encoder_attention_mask=image_attention_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )

        attention_output = attn_outputs[0]

        outputs = attn_outputs[1:]

        layernorm_output = self.post_attention_layernorm(attention_output)

        # Get residual
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = attention_output

        # MLP.
        output = self.gated_mlp(layernorm_output, residual)

        if use_cache:
            outputs = (output,) + outputs
        else:
            outputs = (output,) + outputs[1:]

        return outputs  # hidden_states, present, attentions


class VBloomPreTrainedModel(VLOOMPreTrainedModelBase):
    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = VBloomConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    _no_split_modules = ["BloomBlock"]

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module: nn.Module):
        """Initialize the weights."""
        if isinstance(module, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

    def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False):
        if isinstance(module, VBloomModel):
            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.transformer if hasattr(model, "transformer") else model
        cls.override_vision_model(beheaded_model, vision_model_name, vision_model_params, torch_dtype)
        beheaded_model.freeze_relevant_params(config)


BLOOM_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 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 ([`BloomConfig`]): 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.
"""

BLOOM_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values[0][0].shape[2]`
            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

            [What are input IDs?](../glossary#input-ids)
        past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
            `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
            their past given to this model should not be passed as `input_ids` as they have already been computed.

            Each element of `past_key_values` is a tuple (past_key, past_value):
            - past_key: [batch_size * num_heads, head_dim, kv_length]
            - past_value: [batch_size * num_heads, kv_length, head_dim]
        attention_mask (`torch.FloatTensor` 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.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:

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

        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.

            If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
            `past_key_values`).
        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 [`~file_utils.ModelOutput`] instead of a plain tuple.
"""


@add_start_docstrings(
    "The bare Bloom Model transformer outputting raw hidden-states without any specific head on top.",
    BLOOM_START_DOCSTRING,
)
class VBloomModel(VBloomPreTrainedModel):
    def __init__(self, config: VBloomConfig, vision_model=None):
        super().__init__(config)

        self.embed_dim = config.hidden_size
        self.num_heads = config.n_head

        # Embedding + LN Embedding
        self.word_embeddings = DecoupledEmbedding(
            num_embeddings=config.vocab_size,
            num_additional_embeddings=config.additional_vocab_size,
            embedding_dim=self.embed_dim,
            partially_freeze=config.freeze_text_layers,
        )
        self.word_embeddings_layernorm = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

        # Transformer blocks
        self.h = nn.ModuleList([BloomBlock(config) for _ in range(config.num_hidden_layers)])

        # Final Layer Norm
        self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

        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(
            [VBloomGatedCrossAttentionBlock(config) for i in range(num_cross_layers)]
        )

        # 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,
            )
        self.gradient_checkpointing = False

        # 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

        # 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()

        if config.freeze_vision_layers:
            freeze_model(self.vision_model)

    def freeze_text_layers(self):
        for module in [self.word_embeddings_layernorm, self.h, self.ln_f]:
            freeze_model(module)

    def get_input_embeddings(self):
        return self.word_embeddings

    def _prepare_attn_mask(
        self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
    ) -> torch.BoolTensor:
        # create causal mask
        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
        combined_attention_mask = None
        device = attention_mask.device
        _, src_length = input_shape

        if src_length > 1:
            combined_attention_mask = _make_causal_mask(
                input_shape, device=device, past_key_values_length=past_key_values_length
            )

        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
        expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
        combined_attention_mask = (
            expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
        )

        return combined_attention_mask

    def set_input_embeddings(self, new_embeddings: torch.Tensor):
        self.word_embeddings = new_embeddings

    @add_start_docstrings_to_model_forward(BLOOM_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        processor_class=_TOKENIZER_FOR_DOC,
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=BaseModelOutputWithPastAndCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.LongTensor] = 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,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **deprecated_arguments,
    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
        if deprecated_arguments.pop("position_ids", False) is not False:
            # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
            warnings.warn(
                (
                    "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely"
                    " ignore passing `position_ids`."
                ),
                FutureWarning,
            )
        if len(deprecated_arguments) > 0:
            raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")

        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

        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:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if past_key_values is None:
            past_key_values = tuple([None] * len(self.h))

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape batch_size x num_heads x N x N
        # head_mask has shape n_layer x batch x num_heads x N x N
        head_mask = self.get_head_mask(head_mask, self.config.n_layer)

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

        hidden_states = self.word_embeddings_layernorm(inputs_embeds)

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None

        # Compute alibi tensor: check build_alibi_tensor documentation
        seq_length_with_past = seq_length
        past_key_values_length = 0
        if past_key_values[0] is not None:
            past_key_values_length = past_key_values[0][0].shape[2]
            seq_length_with_past = seq_length_with_past + past_key_values_length
        if attention_mask is None:
            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
        else:
            attention_mask = attention_mask.to(hidden_states.device)

        alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)

        causal_mask = self._prepare_attn_mask(
            attention_mask,
            input_shape=(batch_size, seq_length),
            past_key_values_length=past_key_values_length,
        )

        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
        )  # TODO: something i don't understand here. why are the few last tokens not attending when there is just a single image?
        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=hidden_states.device)
            # image_attention_mask = self.invert_attention_mask(image_attention_mask)
            image_attention_mask = image_attention_mask.to(torch.bool)
            image_attention_mask = image_attention_mask[:, None, :, :]
        else:
            image_attention_mask = None

        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            def vblock(
                main_block,
                hidden_states,
                alibi,
                layer_past,
                attention_mask,
                layer_head_mask,
                use_cache,
                output_attentions,
                image_hidden_states,
                image_attention_mask,
                layer_idx,
                cross_layer_interval,
                gated_cross_attn_layers,
            ):
                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,
                        image_hidden_states=image_hidden_states,
                        image_attention_mask=image_attention_mask,
                        use_cache=use_cache,
                        output_attentions=output_attentions,
                    )
                    hidden_states = outputs[0]

                outputs = main_block(
                    hidden_states,
                    alibi=alibi,
                    layer_past=layer_past,
                    attention_mask=attention_mask,
                    head_mask=layer_head_mask,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

                return outputs

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

                outputs = torch.utils.checkpoint.checkpoint(
                    vblock,
                    block,
                    hidden_states,
                    alibi,
                    layer_past,
                    causal_mask,
                    head_mask[i],
                    use_cache,
                    output_attentions,
                    image_hidden_states,
                    image_attention_mask,
                    i,
                    self.cross_layer_interval,
                    self.gated_cross_attn_layers,
                )
            else:
                outputs = vblock(
                    block,
                    hidden_states,
                    alibi=alibi,
                    layer_past=layer_past,
                    attention_mask=causal_mask,
                    layer_head_mask=head_mask[i],
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                    image_hidden_states=image_hidden_states,
                    image_attention_mask=image_attention_mask,
                    layer_idx=i,
                    cross_layer_interval=self.cross_layer_interval,
                    gated_cross_attn_layers=self.gated_cross_attn_layers,
                )

            hidden_states = outputs[0]
            if use_cache is True:
                presents = presents + (outputs[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)

        # Add last hidden state
        hidden_states = self.ln_f(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


@add_start_docstrings(
    """
    The Bloom Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    BLOOM_START_DOCSTRING,
)
class VBloomForCausalLM(VBloomPreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]

    def __init__(self, config: VBloomConfig, vision_model=None):
        super().__init__(config)
        self.transformer = VBloomModel(config, vision_model=vision_model)
        self.lm_head = DecoupledLinear(
            in_features=config.hidden_size,
            out_features=config.vocab_size,
            out_additional_features=config.additional_vocab_size,
            bias=False,
            partially_freeze=config.freeze_lm_head,
        )
        # Initialize weights and apply final processing
        self.post_init()

    def get_output_embeddings(self):
        return self.lm_head

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

    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 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)

    @add_start_docstrings_to_model_forward(BLOOM_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        processor_class=_TOKENIZER_FOR_DOC,
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=CausalLMOutputWithCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = 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.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **deprecated_arguments,
    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """
        if deprecated_arguments.pop("position_ids", False) is not False:
            # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
            warnings.warn(
                (
                    "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely"
                    " ignore passing `position_ids`."
                ),
                FutureWarning,
            )
        if len(deprecated_arguments) > 0:
            raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")

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

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            head_mask=head_mask,
            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,
        )
        hidden_states = transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)

        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 = lm_logits[..., :-1, :][shift_attention_mask != 0].contiguous()
                shift_labels = labels[..., 1:][shift_attention_mask != 0].contiguous()
            else:
                shift_logits = lm_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 = (lm_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )

    @staticmethod
    def _reorder_cache(
        past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
        """
        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
        beam_idx at every generation step.

        Output shares the same memory storage as `past`.
        """
        batch_size_times_num_heads, head_dim, seq_length = past[0][0].shape
        batch_size = len(beam_idx)
        num_heads = batch_size_times_num_heads // batch_size
        # Get a copy of `beam_idx` on all the devices where we need those indices.
        device_to_beam_idx = {
            past_state.device: beam_idx.to(past_state.device) for layer_past in past for past_state in layer_past
        }
        # key: layer_past[0] [batch_size * num_heads, head_dim, seq_length]
        # value: layer_past[1] [batch_size * num_heads, seq_length, head_dim]
        return tuple(
            (
                layer_past[0]
                .view(batch_size, num_heads, head_dim, seq_length)
                .index_select(0, device_to_beam_idx[layer_past[0].device])
                .view(batch_size_times_num_heads, head_dim, seq_length),
                layer_past[1]
                .view(batch_size, num_heads, seq_length, head_dim)
                .index_select(0, device_to_beam_idx[layer_past[0].device])
                .view(batch_size_times_num_heads, seq_length, head_dim),
            )
            for layer_past in 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
        vision_config = self.transformer.vision_model.config
        num_language_layers = config_vl_model.n_layer
        ffn_inner_size = 4 * config_vl_model.hidden_size

        # 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=False,
            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