# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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 Qwen2 model."""
import inspect
import math
import warnings
from typing import List, Optional, Tuple, Union
import random
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import (
    _prepare_4d_causal_attention_mask,
    _prepare_4d_causal_attention_mask_for_sdpa,
)
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
from typing import List, Optional, Tuple, Union, Dict
import torch
import torch.nn as nn
from transformers import AutoConfig, AutoModelForCausalLM, LlamaConfig, LlamaModel, LlamaForCausalLM, Cache
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.generation.utils import GenerateOutput
from abc import ABC, abstractmethod
import math
import re
import time
import torch
import torch.nn as nn

IGNORE_INDEX = -100
IMAGE_TOKEN_INDEX = -200
DEFAULT_IMAGE_TOKEN = "<image>"
DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
DEFAULT_IM_START_TOKEN = "<im_start>"
DEFAULT_IM_END_TOKEN = "<im_end>"
from typing import Optional, Tuple, Union, Dict
from dataclasses import dataclass
from functools import partial, reduce
from PIL import Image
import torch
import torch.utils.checkpoint
from torch import nn
import os
from transformers.image_processing_utils import BatchFeature, get_size_dict
from transformers.image_transforms import (
    convert_to_rgb,
    normalize,
    rescale,
    resize,
    to_channel_dimension_format,
)
from transformers.image_utils import (
    ChannelDimension,
    PILImageResampling,
    to_numpy_array,
)
from transformers.activations import ACT2FN
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
from transformers.modeling_utils import PreTrainedModel
from transformers import PretrainedConfig
from transformers.utils import ModelOutput
import torch
import torch.nn as nn
import re
import torch.distributed as dist
import math
from .configuration_llavaqwen import LlavaQwenConfig



def rank0_print(*args):
    if dist.is_initialized():
        if dist.get_rank() == 0:
            print(f"Rank {dist.get_rank()}: ", *args)
    else:
        print(*args)


def rank_print(*args):
    if dist.is_initialized():
        print(f"Rank {dist.get_rank()}: ", *args)
    else:
        print(*args)

class PoolerProjector(nn.Module):
    def __init__(self, config, vision_cfg):
        super().__init__()
        self._config = config
        self.hw = vision_cfg.image_size // vision_cfg.patch_size

        self.conv_pool = nn.Conv2d(config.mm_hidden_size, config.hidden_size, kernel_size=2, stride=2)

        self.proj = nn.Sequential(
            nn.GELU(),
            nn.Linear(config.hidden_size, config.hidden_size),
        )

    def forward(self, x, *args, **kwargs):
        height = width = self.hw
        assert height * width == x.shape[1]
        x = x.view(x.shape[0], height, width, -1).permute(0, 3, 1, 2)
        x = self.conv_pool(x)
        x = x.flatten(2).transpose(1, 2)
        x = self.proj(x)
        return x

    @property
    def config(self):
        return {"mm_projector_type": "pooler"}

class IdentityMap(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x, *args, **kwargs):
        return x

    @property
    def config(self):
        return {"mm_projector_type": "identity"}


class SimpleResBlock(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.pre_norm = nn.LayerNorm(channels)

        self.proj = nn.Sequential(nn.Linear(channels, channels), nn.GELU(), nn.Linear(channels, channels))

    def forward(self, x):
        x = self.pre_norm(x)
        return x + self.proj(x)


def build_vision_projector(config, delay_load=False, **kwargs):
    projector_type = getattr(config, "mm_projector_type", "linear")
    mlp_gelu_match = re.match(r"^mlp(\d+)x_gelu$", projector_type)
    if mlp_gelu_match:
        mlp_depth = int(mlp_gelu_match.group(1))
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
        return nn.Sequential(*modules)

class SigLipImageProcessor:
    def __init__(self, image_mean=(0.5, 0.5, 0.5), image_std=(0.5, 0.5, 0.5), size=(384, 384), crop_size: Dict[str, int] = None, resample=PILImageResampling.BICUBIC, rescale_factor=1 / 255, data_format=ChannelDimension.FIRST):
        crop_size = crop_size if crop_size is not None else {"height": 384, "width": 384}
        crop_size = get_size_dict(crop_size, default_to_square=True, param_name="crop_size")

        self.image_mean = image_mean
        self.image_std = image_std
        self.size = size
        self.resample = resample
        self.rescale_factor = rescale_factor
        self.data_format = data_format
        self.crop_size = crop_size

    def preprocess(self, images, return_tensors):
        if isinstance(images, Image.Image):
            images = [images]
        else:
            # to adapt video data
            images = [to_numpy_array(image) for image in images]
            assert isinstance(images, list)

        transforms = [
            convert_to_rgb,
            to_numpy_array,
            partial(resize, size=self.size, resample=self.resample, data_format=self.data_format),
            partial(rescale, scale=self.rescale_factor, data_format=self.data_format),
            partial(normalize, mean=self.image_mean, std=self.image_std, data_format=self.data_format),
            partial(to_channel_dimension_format, channel_dim=self.data_format, input_channel_dim=self.data_format),
        ]

        images = reduce(lambda x, f: [*map(f, x)], transforms, images)
        data = {"pixel_values": images}

        return BatchFeature(data=data, tensor_type=return_tensors)


class SigLipVisionConfig(PretrainedConfig):
    model_type = "siglip_vision_model"

    def __init__(
        self,
        hidden_size=1152,
        image_mean=(0.5, 0.5, 0.5),
        intermediate_size=4304,
        num_hidden_layers=27,
        num_attention_heads=16,
        num_channels=3,
        image_size=384,
        patch_size=14,
        hidden_act="gelu_pytorch_tanh",
        layer_norm_eps=1e-6,
        attention_dropout=0.0,
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_channels = num_channels
        self.patch_size = patch_size
        self.image_size = image_size
        self.attention_dropout = attention_dropout
        self.layer_norm_eps = layer_norm_eps
        self.hidden_act = hidden_act
        self.image_mean = image_mean

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        cls._set_token_in_kwargs(kwargs)

        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)

        # get the vision config dict if we are loading from SigLipConfig
        if config_dict.get("model_type") == "siglip":
            config_dict = config_dict["vision_config"]

        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
            print(f"You are using a model of type {config_dict['model_type']} to instantiate a model of type " f"{cls.model_type}. This is not supported for all configurations of models and can yield errors.")

        return cls.from_dict(config_dict, **kwargs)


@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->SigLip
class SigLipVisionModelOutput(ModelOutput):
    """
    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.

    Args:
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
            The image embeddings obtained by applying the projection layer to the pooler_output.
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

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

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

    image_embeds: Optional[torch.FloatTensor] = None
    last_hidden_state: torch.FloatTensor = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


class SigLipVisionEmbeddings(nn.Module):
    def __init__(self, config: SigLipVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size

        self.patch_embedding = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=self.embed_dim,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            padding="valid",
        )

        self.num_patches = (self.image_size // self.patch_size) ** 2
        self.num_positions = self.num_patches
        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)

    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
        embeddings = patch_embeds.flatten(2).transpose(1, 2)

        embeddings = embeddings + self.position_embedding(self.position_ids)
        return embeddings


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

    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:" f" {self.num_heads}).")
        self.scale = self.head_dim**-0.5
        self.dropout = config.attention_dropout

        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)

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

        batch_size, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        k_v_seq_len = key_states.shape[-2]
        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale

        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
            raise ValueError(f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is" f" {attn_weights.size()}")

        if attention_mask is not None:
            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
                raise ValueError(f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}")
            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
            raise ValueError(f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is" f" {attn_output.size()}")

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->SigLip
class SigLipMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->SigLip
class SigLipEncoderLayer(nn.Module):
    def __init__(self, config: SigLipVisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = SigLipAttention(config)
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = SigLipMLP(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)

    # Ignore copy
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.FloatTensor]:
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                Input to the layer of shape `(batch, seq_len, embed_dim)`.
            attention_mask (`torch.FloatTensor`):
                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class SigLipPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = SigLipVisionConfig
    base_model_prefix = "siglip"
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        """Initialize the weights"""
        pass


# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->SigLip
class SigLipEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`SigLipEncoderLayer`].

    Args:
        config: SigLipVisionConfig
    """

    def __init__(self, config: SigLipVisionConfig):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([SigLipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    # Ignore copy
    def forward(
        self,
        inputs_embeds,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                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.
            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)
            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.
        """
        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

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    encoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    output_attentions,
                )
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask,
                    output_attentions=output_attentions,
                )

            hidden_states = layer_outputs[0]

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

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions)


class SigLipVisionTransformer(nn.Module):
    def __init__(self, config: SigLipVisionConfig):
        super().__init__()
        self.config = config
        embed_dim = config.hidden_size

        self.embeddings = SigLipVisionEmbeddings(config)
        self.encoder = SigLipEncoder(config)
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
        self.head = SigLipMultiheadAttentionPoolingHead(config)

    def forward(
        self,
        pixel_values,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        r"""
        Returns:

        """
        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

        hidden_states = self.embeddings(pixel_values)

        encoder_outputs = self.encoder(
            inputs_embeds=hidden_states,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        last_hidden_state = encoder_outputs[0]
        last_hidden_state = self.post_layernorm(last_hidden_state)

        pooled_output = self.head(last_hidden_state)

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

        return BaseModelOutputWithPooling(
            last_hidden_state=last_hidden_state,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


class SigLipMultiheadAttentionPoolingHead(nn.Module):
    """Multihead Attention Pooling."""

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

        self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
        self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.mlp = SigLipMLP(config)

    def forward(self, hidden_state):
        batch_size = hidden_state.shape[0]
        probe = self.probe.repeat(batch_size, 1, 1)

        hidden_state = self.attention(probe, hidden_state, hidden_state)[0]

        residual = hidden_state
        hidden_state = self.layernorm(hidden_state)
        hidden_state = residual + self.mlp(hidden_state)

        return hidden_state[:, 0]


class SigLipVisionModel(SigLipPreTrainedModel):
    config_class = SigLipVisionConfig
    main_input_name = "pixel_values"
    _no_split_modules = ["SigLipEncoderLayer"]

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

        self.vision_model = SigLipVisionTransformer(config)

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

    def get_input_embeddings(self) -> nn.Module:
        return self.vision_model.embeddings.patch_embedding

    def forward(
        self,
        pixel_values,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        r"""
        Returns:

        Examples:

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

        >>> model = SigLipVisionModel.from_pretrained("google/siglip-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

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

        >>> outputs = model(**inputs)
        >>> last_hidden_state = outputs.last_hidden_state
        >>> pooled_output = outputs.pooler_output  # pooled features
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        return self.vision_model(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )


class SigLipVisionTower(nn.Module):
    def __init__(self, vision_tower, vision_tower_cfg, delay_load=False):
        super().__init__()

        self.is_loaded = False

        self.config = SigLipVisionConfig()

        self.vision_tower_name = vision_tower

        self.image_processor = SigLipImageProcessor()

        if not delay_load:
            rank0_print(f"Loading vision tower: {vision_tower}")
            self.load_model()
        elif getattr(vision_tower_cfg, "unfreeze_mm_vision_tower", False):
            # TODO: better detector is needed.
            rank0_print(f"The checkpoint seems to contain `vision_tower` weights: `unfreeze_mm_vision_tower`: True.")
            self.load_model()
        elif hasattr(vision_tower_cfg, "mm_tunable_parts") and "mm_vision_tower" in vision_tower_cfg.mm_tunable_parts:
            rank0_print(f"The checkpoint seems to contain `vision_tower` weights: `mm_tunable_parts` contains `mm_vision_tower`.")
            self.load_model()
        else:
            self.cfg_only = self.config

    def load_model(self, device_map=None):
        if self.is_loaded:
            rank0_print("{} is already loaded, `load_model` called again, skipping.".format(self.vision_tower_name))
            return

        self.vision_tower = SigLipVisionModel.from_pretrained(self.vision_tower_name, device_map=device_map)

        del self.vision_tower.vision_model.encoder.layers[-1:]
        self.vision_tower.vision_model.head = nn.Identity()
        self.vision_tower.requires_grad_(False)

        self.is_loaded = True

    def forward(self, images):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = image_forward_out.hidden_states[-1].to(image.dtype)
                assert image_features.shape[-2] == 729
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
            image_features = image_forward_outs.hidden_states[-1].to(images.dtype)
            assert image_features.shape[-2] == 729

        return image_features

    @property
    def dummy_feature(self):
        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)

    @property
    def dtype(self):
        for p in self.vision_tower.parameters():
            return p.dtype

    @property
    def device(self):
        for p in self.vision_tower.parameters():
            return p.device

    @property
    def hidden_size(self):
        return self.config.hidden_size

    @property
    def num_patches(self):
        return (self.config.image_size // self.config.patch_size) ** 2

    @property
    def num_patches_per_side(self):
        return self.config.image_size // self.config.patch_size
        # return self.model_config["vision_cfg"]["image_size"] // self.model_config["vision_cfg"]["patch_size"]

    @property
    def image_size(self):
        return self.config.image_size


def build_vision_tower(vision_tower_cfg, **kwargs):
    vision_tower = getattr(vision_tower_cfg, "mm_vision_tower", getattr(vision_tower_cfg, "vision_tower", None))
    is_absolute_path_exists = os.path.exists(vision_tower)
    use_s2 = getattr(vision_tower_cfg, "s2", False)
    return SigLipVisionTower(vision_tower, vision_tower_cfg=vision_tower_cfg, **kwargs)


class IdentityMap(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x, *args, **kwargs):
        return x

    @property
    def config(self):
        return {"mm_resampler_type": None}


def build_vision_resampler(model_args, delay_load=False, **kwargs):
    resampler_type = getattr(model_args, "mm_resampler_type", None)
    return IdentityMap()


class LlavaMetaModel:

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

        if hasattr(config, "mm_vision_tower"):
            delay_load = getattr(config, "delay_load", False)
            self.vision_tower = build_vision_tower(config, delay_load=delay_load)
            self.vision_resampler = build_vision_resampler(config, vision_tower=self.vision_tower)
            self.mm_projector = build_vision_projector(config, vision_cfg=self.vision_tower.config)

            if "unpad" in getattr(config, "mm_patch_merge_type", ""):
                self.image_newline = nn.Parameter(torch.empty(config.hidden_size, dtype=self.dtype))

    def get_vision_tower(self):
        vision_tower = getattr(self, "vision_tower", None)
        if type(vision_tower) is list:
            vision_tower = vision_tower[0]
        return vision_tower

    def initialize_vision_modules(self, model_args, fsdp=None):
        vision_tower = model_args.vision_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
        mm_patch_merge_type = model_args.mm_patch_merge_type

        self.config.mm_vision_tower = vision_tower
        self.config.vision_tower_pretrained = getattr(model_args, "vision_tower_pretrained", "")

        if self.get_vision_tower() is None:
            vision_tower = build_vision_tower(model_args)
            vision_resampler = build_vision_resampler(model_args, vision_tower=vision_tower)
            for k, v in vision_resampler.config.items():
                setattr(self.config, k, v)

            if fsdp is not None and len(fsdp) > 0:
                self.vision_tower = [vision_tower]
                self.vision_resampler = [vision_resampler]
            else:
                self.vision_tower = vision_tower
                self.vision_resampler = vision_resampler
        else:
            if fsdp is not None and len(fsdp) > 0:
                vision_resampler = self.vision_resampler[0]
                vision_tower = self.vision_tower[0]
            else:
                vision_resampler = self.vision_resampler
                vision_tower = self.vision_tower
            vision_tower.load_model()

            # In case it is frozen by LoRA
            for p in self.vision_resampler.parameters():
                p.requires_grad = True

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, "mm_projector_type", "linear")
        self.config.mm_hidden_size = getattr(vision_resampler, "hidden_size", vision_tower.hidden_size)
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature
        self.config.mm_patch_merge_type = mm_patch_merge_type

        
        if not hasattr(self.config, 'add_faster_video'):
            if model_args.add_faster_video:
                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
                self.faster_token = nn.Parameter(
                    torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std
                )

        if getattr(self, "mm_projector", None) is None:
            self.mm_projector = build_vision_projector(self.config, vision_cfg=vision_tower.config)

            if "unpad" in mm_patch_merge_type:
                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
                self.image_newline = nn.Parameter(torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std)
        else:
            # In case it is frozen by LoRA
            for p in self.mm_projector.parameters():
                p.requires_grad = True

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location="cpu")

            def get_w(weights, keyword):
                return {k.split(keyword + ".")[1]: v for k, v in weights.items() if keyword in k}

            incompatible_keys = self.mm_projector.load_state_dict(get_w(mm_projector_weights, "mm_projector"))
            rank0_print(f"Loaded mm projector weights from {pretrain_mm_mlp_adapter}. Incompatible keys: {incompatible_keys}")
            incompatible_keys = self.vision_resampler.load_state_dict(get_w(mm_projector_weights, "vision_resampler"), strict=False)
            rank0_print(f"Loaded vision resampler weights from {pretrain_mm_mlp_adapter}. Incompatible keys: {incompatible_keys}")


def unpad_image(tensor, original_size):
    """
    Unpads a PyTorch tensor of a padded and resized image.

    Args:
    tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
    original_size (tuple): The original size of the image (height, width).

    Returns:
    torch.Tensor: The unpadded image tensor.
    """
    original_width, original_height = original_size
    current_height, current_width = tensor.shape[1:]

    # Compute aspect ratios
    original_aspect_ratio = original_width / original_height
    current_aspect_ratio = current_width / current_height

    # Determine padding size and direction
    if original_aspect_ratio > current_aspect_ratio:
        # Padding was added to the height
        scale_factor = current_width / original_width
        new_height = int(original_height * scale_factor)
        padding = (current_height - new_height) // 2
        unpadded_tensor = tensor[:, padding : current_height - padding, :]
    else:
        # Padding was added to the width
        scale_factor = current_height / original_height
        new_width = int(original_width * scale_factor)
        padding = (current_width - new_width) // 2
        unpadded_tensor = tensor[:, :, padding : current_width - padding]

    return unpadded_tensor


class LlavaMetaForCausalLM(ABC):

    @abstractmethod
    def get_model(self):
        pass

    def get_vision_tower(self):
        return self.get_model().get_vision_tower()

    def get_2dPool(self, image_feature, stride=2):
        height = width = self.get_vision_tower().num_patches_per_side
        num_frames, num_tokens, num_dim = image_feature.shape
        image_feature = image_feature.view(num_frames, height, width, -1)
        image_feature = image_feature.permute(0, 3, 1, 2).contiguous()
        # image_feature = nn.functional.max_pool2d(image_feature, self.config.mm_spatial_pool_stride)
        if self.config.mm_spatial_pool_mode == "average":
            image_feature = nn.functional.avg_pool2d(image_feature, stride)
        elif self.config.mm_spatial_pool_mode == "max":
            image_feature = nn.functional.max_pool2d(image_feature, stride)
        elif self.config.mm_spatial_pool_mode == "bilinear":
            height, width = image_feature.shape[2:]
            scaled_shape = [math.ceil(height / stride), math.ceil(width / stride)]
            image_feature = nn.functional.interpolate(image_feature, size=scaled_shape, mode='bilinear')

        else:
            raise ValueError(f"Unexpected mm_spatial_pool_mode: {self.config.mm_spatial_pool_mode}")
        image_feature = image_feature.permute(0, 2, 3, 1)
        image_feature = image_feature.view(num_frames, -1, num_dim)
        return image_feature

    def encode_multimodals(self, videos_or_images, video_idx_in_batch, split_sizes=None):
        videos_or_images_features = self.get_model().get_vision_tower()(videos_or_images)
        per_videos_or_images_features = torch.split(videos_or_images_features, split_sizes, dim=0)  # tuple, (dim_1, 576, 4096)
        all_videos_or_images_features = []
        all_faster_video_features = []
        cur_mm_spatial_pool_stride = self.config.mm_spatial_pool_stride

        for idx, feat in enumerate(per_videos_or_images_features):
            
            feat = self.get_model().mm_projector(feat)
            faster_video_feature = 0
            slower_img_feat = 0
            if idx in video_idx_in_batch and cur_mm_spatial_pool_stride > 1:
                slower_img_feat = self.get_2dPool(feat,cur_mm_spatial_pool_stride)
                if self.config.add_faster_video:
                    cur_mm_spatial_pool_stride = cur_mm_spatial_pool_stride * 2
                    faster_video_feature = self.get_2dPool(feat,cur_mm_spatial_pool_stride)
            if slower_img_feat is not 0:
                all_videos_or_images_features.append(slower_img_feat)
            else:
                all_videos_or_images_features.append(feat)
            all_faster_video_features.append(faster_video_feature)
        return all_videos_or_images_features,all_faster_video_features

    def add_token_per_grid(self, image_feature):
        resize_h = int(math.sqrt(image_feature.shape[1]))
        num_frames = image_feature.shape[0]
        feature_dim = image_feature.shape[-1]

        image_feature = image_feature.view(num_frames, 1, resize_h, resize_h, -1)
        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
        image_feature = torch.cat((image_feature, self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)), dim=-1)
        if getattr(self.config, "add_faster_video", False):
            # import pdb; pdb.set_trace()
            # (3584, 832, 14) -> (3584, 64, 13, 14)
            image_feature = image_feature.view(feature_dim, num_frames,resize_h, -1)
            #  (3584, 64, 13, 14) -> (64, 13, 14, 3584)
            image_feature = image_feature.permute(1, 2, 3, 0).contiguous()
            # (64, 13, 14, 3584) -> (64, 13*14, 3584)
            image_feature = image_feature.flatten(1, 2)
            # import pdb; pdb.set_trace()
            return image_feature
        # import pdb; pdb.set_trace()
        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
        return image_feature

    def add_token_per_frame(self, image_feature):
        image_feature = image_feature.permute(2, 0, 1).contiguous()
        image_feature =  torch.cat((image_feature, self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)), dim=-1)
        image_feature = image_feature.permute(1, 2, 0).contiguous()
        return image_feature

    def prepare_inputs_labels_for_multimodal_interleave(
            self,
            input_ids,
            position_ids,
            attention_mask,
            past_key_values,
            labels,
            images,
            modalities,
            clip_sizes,
            image_sizes_per_clip,
            prompts=None,
    ):
        vision_tower = self.get_vision_tower()
        if vision_tower is None or images is None or input_ids.shape[1] == 1:
            return input_ids, position_ids, attention_mask, past_key_values, None, labels

        # print('clip_sizes', clip_sizes)
        # print('image_sizes_per_clip', image_sizes_per_clip)
        # print('images', images.shape)

        # breakpoint()
        if isinstance(modalities, str):
            modalities = [modalities]

        if torch.cuda.current_device() == 0:
            print(f'[RANK0 PRINT] | Modality Check: {modalities}')

        if type(images) is list or images.ndim == 5:
            if type(images) is list:
                # batch of list of images, B x [N x C x H x W]
                images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]

            video_idx_in_batch = []
            for _ in range(len(modalities)):
                if modalities[_] in ["video"]:
                    video_idx_in_batch.append(_)

            images_list = []
            for image in images:
                if image.ndim == 4:
                    images_list.append(image)
                else:
                    images_list.append(image.unsqueeze(0))

            concat_images = torch.cat([image for image in images_list], dim=0)
            # this records num_frames for each sample
            split_sizes = [image.shape[0] for image in images_list]
            # list of video-level image features: B x [N x P x D]
            image_features = self.encode_images(concat_images, video_idx_in_batch, split_sizes)

            # below this line, we switch the process unit from video to clip
            clip_image_features = []
            image_sizes = []
            clip_modalities = []
            for image_idx, image_feature in enumerate(image_features):
                num_frames = image_feature.shape[0]
                clip_size = clip_sizes[image_idx]
                assert sum(clip_size) == num_frames, 'num_frame of image_feature does not match metadata'
                modality = modalities[image_idx]
                image_size = image_sizes_per_clip[image_idx]
                per_clip_features = torch.split(image_feature, clip_size, dim=0)

                clip_image_features.extend(per_clip_features)
                image_sizes.extend(image_size)
                clip_modalities.extend([modality for _ in range(len(clip_size))])

            image_features = clip_image_features

            video_idx_in_batch = []
            for _ in range(len(clip_modalities)):
                if clip_modalities[_] in ["video"]:
                    video_idx_in_batch.append(_)

            # image_features = torch.split(image_features, split_sizes, dim=0)
            mm_patch_merge_type = getattr(self.config, "mm_patch_merge_type", "flat")
            image_aspect_ratio = getattr(self.config, "image_aspect_ratio", "square")
            new_image_features = []

            if mm_patch_merge_type == 'flat':
                for image_idx, image_feature in enumerate(image_features):
                    new_image_features.append(image_feature.flatten(0, 1))
                image_features = new_image_features

            elif mm_patch_merge_type.startswith('spatial'):
                for image_idx, image_feature in enumerate(image_features):
                    # FIXME: now assume the image is square, and split to 2x2 patches
                    # num_patches = h * w, where h = w = sqrt(num_patches)
                    # currently image_feature is a tensor of shape (4, num_patches, hidden_size)
                    # we want to first unflatten it to (2, 2, h, w, hidden_size)

                    if image_feature.shape[0] > 1:
                        if image_idx in video_idx_in_batch:
                            if self.config.mm_newline_position == "grid":
                                # Grid-wise
                                resize_h = int(math.sqrt(image_feature.shape[1]))
                                num_frames = image_feature.shape[0]
                                image_feature = image_feature.view(num_frames, 1, resize_h, resize_h, -1)
                                # N x 1 x H x W x D -> D x N x H x 1 x W
                                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                                # D x N x H x 1 x W -> D x N*H x 1 x W -> D x N*H x W
                                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                                # D x N*H x (W+1)
                                image_feature = torch.cat(
                                    (
                                        image_feature,  # D x N*H x W
                                        self.model.image_newline[:, None, None]  # D x 1 x 1 -> D x N*H x 1
                                        .expand(*image_feature.shape[:-1], 1)
                                        .to(image_feature.device),
                                    ),
                                    dim=-1,
                                )
                                # D x N*H x (W+1) -> D x N*H*(W+1) -> N*H*(W+1) x D
                                image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                                new_image_features.append(image_feature)
                            elif self.config.mm_newline_position == "frame":
                                # Frame-wise
                                image_feature = image_feature.permute(2, 0, 1).contiguous()
                                image_feature = torch.cat(
                                    (
                                        image_feature,
                                        self.model.image_newline[:, None, None]
                                        .expand(*image_feature.shape[:-1], 1)
                                        .to(image_feature.device),
                                    ),
                                    dim=-1,
                                )
                                image_feature = image_feature.permute(1, 2, 0).contiguous()
                                new_image_features.append(image_feature.flatten(0, 1))
                            elif self.config.mm_newline_position == "one_token":
                                # one-token
                                image_feature = image_feature.flatten(0, 1)
                                if 'unpad' in mm_patch_merge_type:
                                    image_feature = torch.cat(
                                        (image_feature, self.model.image_newline[None].to(image_feature.device)), dim=0
                                    )
                                new_image_features.append(image_feature)
                            elif self.config.mm_newline_position == "no_token":
                                new_image_features.append(image_feature.flatten(0, 1))
                            else:
                                raise ValueError(f"Unexpected mm_newline_position: {self.config.mm_newline_position}")

                            continue

                        # 1 x D
                        base_image_feature = image_feature[0]
                        # N*H*W x D
                        image_feature = image_feature[1:]
                        height = width = self.get_vision_tower().num_patches_per_side
                        assert height * width == base_image_feature.shape[0]

                        if "anyres_max" in image_aspect_ratio:
                            matched_anyres_max_num_patches = re.match(r"anyres_max_(\d+)", image_aspect_ratio)
                            if matched_anyres_max_num_patches:
                                max_num_patches = int(matched_anyres_max_num_patches.group(1))
                        if image_aspect_ratio == 'anyres' or "anyres_max" in image_aspect_ratio:
                            try:
                                num_patch_width, num_patch_height = get_anyres_image_grid_shape(
                                    image_sizes[image_idx],
                                    self.config.image_grid_pinpoints,
                                    self.get_vision_tower().config.image_size,
                                )
                            except:
                                raise ValueError("get anyres image grid shape error")
                            # N*H*W*D -> p_H x p_W x H x W x D
                            image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                        else:
                            image_feature = image_feature.view(2, 2, height, width, -1)
                        if 'maxpool2x2' in mm_patch_merge_type:
                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                            image_feature = nn.functional.max_pool2d(image_feature, 2)
                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                        elif (
                                "unpad" in mm_patch_merge_type
                                and "anyres_max" in image_aspect_ratio
                                and matched_anyres_max_num_patches
                        ):
                            unit = image_feature.shape[2]
                            # p_H x p_W x H x W x D -> D x p_H x H x p_W x W
                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                            # D x p_H x H x p_W x W -> D x p_H*H x p_W x W -> D x p_H*H x p_W*W
                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
                            c, h, w = image_feature.shape
                            times = math.sqrt(h * w / (max_num_patches * unit ** 2))
                            if times > 1.1:
                                image_feature = image_feature[None]
                                image_feature = nn.functional.interpolate(
                                    image_feature, [int(h // times), int(w // times)], mode="bilinear"
                                )[0]
                            image_feature = torch.cat(
                                (
                                    image_feature,
                                    self.model.image_newline[:, None, None]
                                    .expand(*image_feature.shape[:-1], 1)
                                    .to(image_feature.device),
                                ),
                                dim=-1,
                            )
                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                        elif 'unpad' in mm_patch_merge_type:
                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
                            image_feature = torch.cat(
                                (
                                    image_feature,
                                    self.model.image_newline[:, None, None]
                                    .expand(*image_feature.shape[:-1], 1)
                                    .to(image_feature.device),
                                ),
                                dim=-1,
                            )
                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                        else:
                            image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
                            image_feature = image_feature.flatten(0, 3)
                        if 'nobase' in mm_patch_merge_type:
                            pass
                        else:
                            image_feature = torch.cat((base_image_feature, image_feature), dim=0)
                    else:
                        image_feature = image_feature[0]
                        if 'unpad' in mm_patch_merge_type:
                            image_feature = torch.cat(
                                (image_feature, self.model.image_newline[None].to(image_feature.device)), dim=0
                            )

                    new_image_features.append(image_feature)
                image_features = new_image_features
            else:
                raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")
        else:
            image_features = self.encode_images(images)

        # TODO: image start / end is not implemented here to support pretraining.
        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
            raise NotImplementedError

        # Let's just add dummy tensors if they do not exist,
        # it is a headache to deal with None all the time.
        # But it is not ideal, and if you have a better idea,
        # please open an issue / submit a PR, thanks.
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, IGNORE_INDEX)

        # remove the padding using attention_mask -- FIXME
        input_ids = [
            cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)
        ]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]

        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0

        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
            if num_images == 0:
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue

            image_token_indices = (
                    [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
            )
            cur_input_ids_noim = []
            cur_labels = labels[batch_idx]
            cur_labels_noim = []
            for i in range(len(image_token_indices) - 1):
                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i] + 1: image_token_indices[i + 1]])
                cur_labels_noim.append(cur_labels[image_token_indices[i] + 1: image_token_indices[i + 1]])
            # get the length of each text
            split_sizes = [x.shape[0] for x in cur_labels_noim]
            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
            cur_new_input_embeds = []
            cur_new_labels = []

            for i in range(num_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
                cur_new_labels.append(cur_labels_noim[i])
                if i < num_images:
                    cur_image_features = image_features[cur_image_idx]
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(
                        torch.full(
                            (cur_image_features.shape[0],),
                            IGNORE_INDEX,
                            device=cur_labels.device,
                            dtype=cur_labels.dtype,
                        )
                    )

            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]

            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_new_labels)

        assert cur_image_idx == len(new_image_features), \
            'not all clip features are inserted, please check input sequence.'

        # Truncate sequences to max length as image embeddings can make the sequence longer
        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
        modality_max_length = getattr(self.config, 'modality_max_length', None)

        if modality_max_length is None or modality_max_length == "None":
            if tokenizer_model_max_length is not None:
                new_input_embeds = [x[:tokenizer_model_max_length] for x, modality in zip(new_input_embeds, modalities)]
                new_labels = [x[:tokenizer_model_max_length] for x, modality in zip(new_labels, modalities)]
        else:
            modality_max_length = ast.literal_eval(modality_max_length)
            modality_max_length_dict = {
                "image": modality_max_length[0],
                "text": modality_max_length[1],
                "video": modality_max_length[2],
            }
            new_input_embeds = [
                x[: modality_max_length_dict[modality]] for x, modality in zip(new_input_embeds, modalities)
            ]
            new_labels = [x[: modality_max_length_dict[modality]] for x, modality in zip(new_labels, modalities)]

        # Combine them
        max_len = max(x.shape[0] for x in new_input_embeds)
        batch_size = len(new_input_embeds)

        new_input_embeds_padded = []
        new_labels_padded = torch.full(
            (batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device
        )
        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)

        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
            cur_len = cur_new_embed.shape[0]
            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                new_input_embeds_padded.append(
                    torch.cat(
                        (
                            torch.zeros(
                                (max_len - cur_len, cur_new_embed.shape[1]),
                                dtype=cur_new_embed.dtype,
                                device=cur_new_embed.device,
                            ),
                            cur_new_embed,
                        ),
                        dim=0,
                    )
                )
                if cur_len > 0:
                    new_labels_padded[i, -cur_len:] = cur_new_labels
                    attention_mask[i, -cur_len:] = True
                    position_ids[i, -cur_len:] = torch.arange(
                        0, cur_len, dtype=position_ids.dtype, device=position_ids.device
                    )
            else:
                new_input_embeds_padded.append(
                    torch.cat(
                        (
                            cur_new_embed,
                            torch.zeros(
                                (max_len - cur_len, cur_new_embed.shape[1]),
                                dtype=cur_new_embed.dtype,
                                device=cur_new_embed.device,
                            ),
                        ),
                        dim=0,
                    )
                )
                if cur_len > 0:
                    new_labels_padded[i, :cur_len] = cur_new_labels
                    attention_mask[i, :cur_len] = True
                    position_ids[i, :cur_len] = torch.arange(
                        0, cur_len, dtype=position_ids.dtype, device=position_ids.device
                    )

        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)

        if _labels is None:
            new_labels = None
        else:
            new_labels = new_labels_padded

        if _attention_mask is None:
            attention_mask = None
        else:
            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

        if _position_ids is None:
            position_ids = None

        if torch.cuda.current_device() == 0:
            print(f'[RANK0 PRINT] | new_input_embeds\'s shape: {new_input_embeds.shape}')

        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels

    def initialize_vision_tokenizer(self, model_args, tokenizer):
        if model_args.mm_use_im_patch_token:
            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

        if model_args.mm_use_im_start_end:
            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

            if num_new_tokens > 0:
                input_embeddings = self.get_input_embeddings().weight.data
                output_embeddings = self.get_output_embeddings().weight.data

                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)

                input_embeddings[-num_new_tokens:] = input_embeddings_avg
                output_embeddings[-num_new_tokens:] = output_embeddings_avg

            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = True
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

            if model_args.pretrain_mm_mlp_adapter:
                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location="cpu")
                embed_tokens_weight = mm_projector_weights["model.embed_tokens.weight"]
                assert num_new_tokens == 2
                if input_embeddings.shape == embed_tokens_weight.shape:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
                elif embed_tokens_weight.shape[0] == num_new_tokens:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
                else:
                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
        elif model_args.mm_use_im_patch_token:
            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = False
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

    def process_images_in_batches(self, model, images, img_bs):
        all_features = []
        total_batches = (len(images) + img_bs - 1) // img_bs

        for i in range(total_batches):
            batch_images = images[i * img_bs: (i + 1) * img_bs]
            features = model(batch_images)
            all_features.append(features)
        all_features = torch.cat(all_features, dim=0)

        return all_features

    def encode_images(self, images, video_idx_in_batch=[], split_sizes=None):

        # get vision tower
        vision_tower = self.get_model().get_vision_tower()
        image_features = self.process_images_in_batches(  # NOTE: Hard Code, set max img_bs to 300
            vision_tower, images, img_bs=300
        )  # forward, (num_images, 3, 384, 384) -> (num_images, 729, 1152)

        # [opt] get frame selector
        smarter_frame = False
        if getattr(self.config, 'mm_smarter_frames_sel_strategy', "all") == "gate_fix":
            frame_selector = self.get_model().get_frame_selector()
            smarter_frame = True

        if split_sizes is None:
            split_sizes = [1 for image in images]
        # split images according to each sample's num_frames, i.e., split_sizes
        per_image_features = torch.split(image_features, split_sizes, dim=0)  # tuple, (num_images, 729, 1152)
        all_image_features = []

        # breakpoint()
        for idx, img_feat in enumerate(per_image_features):
            # img_feat: (num_images, 729, 1152)

            # frame seletcion: (num_images, 729, 1152) -> (num_images', 1) -> (topk_images, 729, 1152)
            if smarter_frame and (self.config.mm_smarter_frames_sel_position == "before"):
                img_feat = frame_selector(img_feat)

            # patch pooling
            if self.config.mm_pooling_position == "before":
                if idx in video_idx_in_batch and self.config.mm_spatial_pool_stride > 1:
                    img_feat = self.get_2dPool(img_feat)  # (num_images, 169, 1152) -> (num_images, 169, 1152)

            # Projector here!!!
            img_feat = self.get_model().mm_projector(img_feat)  # (num_images, 169, 1152) -> (num_images, 169, 3584)

            # frame seletcion: (num_images, 169, 3584) -> (num_images', 1) -> (topk_images, 169, 3584)
            if smarter_frame and (self.config.mm_smarter_frames_sel_position == "after"):
                img_feat = frame_selector(img_feat)

            # patch pooling
            if self.config.mm_pooling_position == "after":
                if idx in video_idx_in_batch and self.config.mm_spatial_pool_stride > 1:
                    img_feat = self.get_2dPool(img_feat)  # (num_images, 169, 3584) -> (num_images, 169, 3584)

            all_image_features.append(img_feat)

        return all_image_features


faster_llama_rmsnorm = None
if is_flash_attn_2_available():
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa


    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)

logger = logging.get_logger(__name__)

try:
    from flash_attn.layers.rotary import apply_rotary_emb_func
except:
    apply_rotary_emb_func = None
    logger.warning_once('fail to load faster rotary ops, use PyTorch version by default. Please check image version')

_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta"
_CONFIG_FOR_DOC = "Qwen2Config"

QWEN2_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "Qwen/Qwen2-7B-beta",
    # See all Qwen2 models at https://huggingface.co/models?filter=qwen2
]


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
class Qwen2RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Qwen2RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        if faster_llama_rmsnorm:
            if not isinstance(self.variance_epsilon, torch.Tensor):
                self.variance_epsilon = torch.tensor(
                    self.variance_epsilon, dtype=self.weight.dtype, device=self.weight.device
                )
            if len(hidden_states.shape) == 2:
                hidden_states = hidden_states.view(1, hidden_states.shape[0], hidden_states.shape[1])
                return faster_llama_rmsnorm(hidden_states, self.weight, self.variance_epsilon).squeeze(0)
            else:
                return faster_llama_rmsnorm(hidden_states, self.weight, self.variance_epsilon)
        else:
            input_dtype = hidden_states.dtype
            hidden_states = hidden_states.to(torch.float32)
            variance = hidden_states.pow(2).mean(-1, keepdim=True)
            hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
            return self.weight * hidden_states.to(input_dtype)


# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Qwen2
class Qwen2RotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
        super().__init__()

        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        # Build here to make `torch.jit.trace` work.
        self._set_cos_sin_cache(
            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
        )

    def _set_cos_sin_cache(self, seq_len, device, dtype):
        self.max_seq_len_cached = seq_len
        t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)

        freqs = torch.outer(t, self.inv_freq)
        # Different from paper, but it uses a different permutation in order to obtain the same calculation
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)

    def forward(self, x, seq_len=None):
        # x: [bs, num_attention_heads, seq_len, head_size]
        if seq_len > self.max_seq_len_cached:
            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)

        return (
            self.cos_cached[:seq_len].to(dtype=x.dtype),
            self.sin_cached[:seq_len].to(dtype=x.dtype),
        )


# Copied from transformers.models.llama.modeling_llama.rotate_half
def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2:]
    return torch.cat((-x2, x1), dim=-1)


# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2
class Qwen2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))


# Copied from transformers.models.llama.modeling_llama.repeat_kv
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


class Qwen2Attention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.attention_dropout = config.attention_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}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.rotary_emb = Qwen2RotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta,
        )

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
            **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class Qwen2FlashAttention2(Qwen2Attention):
    """
    Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
    as the weights of the module stays untouched. The only required change would be on the forward pass
    where it needs to correctly call the public API of flash attention and deal with padding tokens
    in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
    config.max_window_layers layers.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
            **kwargs,
    ):
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            # overwrite attention_mask with padding_mask
            attention_mask = kwargs.pop("padding_mask")
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # Because the input can be padded, the absolute sequence length depends on the max position id.
        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
        cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)

        if apply_rotary_emb_func is not None:
            cos = cos.squeeze().index_select(dim=0, index=position_ids.squeeze())  # [total_bs_seq, head_dim]
            sin = sin.squeeze().index_select(dim=0, index=position_ids.squeeze())
            query_states = apply_rotary_emb_func(
                query_states.transpose(1, 2), cos[:, : self.head_dim // 2], sin[:, : self.head_dim // 2]
            ).transpose(1, 2)
            key_states = apply_rotary_emb_func(
                key_states.transpose(1, 2), cos[:, : self.head_dim // 2], sin[:, : self.head_dim // 2]
            ).transpose(1, 2)
        else:
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

        use_sliding_windows = (
                _flash_supports_window_size
                and getattr(self.config, "sliding_window", None) is not None
                and kv_seq_len > self.config.sliding_window
                and self.config.use_sliding_window
        )

        if not _flash_supports_window_size:
            logger.warning_once(
                "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
                " make sure to upgrade flash-attn library."
            )

        if past_key_value is not None:
            # Activate slicing cache only if the config has a value `sliding_windows` attribute
            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
            if (
                    getattr(self.config, "sliding_window", None) is not None
                    and kv_seq_len > self.config.sliding_window
                    and cache_has_contents
            ):
                slicing_tokens = 1 - self.config.sliding_window

                past_key = past_key_value[self.layer_idx][0]
                past_value = past_key_value[self.layer_idx][1]

                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                past_value = past_value[:, :, slicing_tokens:, :].contiguous()

                if past_key.shape[-2] != self.config.sliding_window - 1:
                    raise ValueError(
                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
                        f" {past_key.shape}"
                    )

                if attention_mask is not None:
                    attention_mask = attention_mask[:, slicing_tokens:]
                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)

            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        dropout_rate = 0.0 if not self.training else self.attention_dropout

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in float16 just to be sure everything works as expected.
        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        # Reashape to the expected shape for Flash Attention
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            use_sliding_windows=use_sliding_windows,
        )

        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

    def _flash_attention_forward(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            query_length,
            dropout=0.0,
            softmax_scale=None,
            use_sliding_windows=False,
    ):
        """
        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
        first unpad the input, then computes the attention scores and pad the final attention scores.

        Args:
            query_states (`torch.Tensor`):
                Input query states to be passed to Flash Attention API
            key_states (`torch.Tensor`):
                Input key states to be passed to Flash Attention API
            value_states (`torch.Tensor`):
                Input value states to be passed to Flash Attention API
            attention_mask (`torch.Tensor`):
                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                position of padding tokens and 1 for the position of non-padding tokens.
            dropout (`float`):
                Attention dropout
            softmax_scale (`float`, *optional*):
                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
            use_sliding_windows (`bool`, *optional*):
                Whether to activate sliding window attention.
        """
        if not self._flash_attn_uses_top_left_mask:
            causal = self.is_causal
        else:
            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
            causal = self.is_causal and query_length != 1

        # Decide whether to use SWA or not by layer index.
        if use_sliding_windows and self.layer_idx >= self.config.max_window_layers:
            use_sliding_windows = False

        # Contains at least one padding token in the sequence
        if attention_mask is not None:
            batch_size = query_states.shape[0]
            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                query_states, key_states, value_states, attention_mask, query_length
            )

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

            if not use_sliding_windows:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            if not use_sliding_windows:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

        return attn_output

    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape

        # On the first iteration we need to properly re-create the padding mask
        # by slicing it on the proper place
        if kv_seq_len != attention_mask.shape[-1]:
            attention_mask_num_tokens = attention_mask.shape[-1]
            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len:]

        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)

        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)

        if query_length == kv_seq_len:
            query_layer = index_first_axis(query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_in_batch_q = max_seqlen_in_batch_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_in_batch_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query_layer.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query_layer = query_layer.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

        return (
            query_layer,
            key_layer,
            value_layer,
            indices_q,
            (cu_seqlens_q, cu_seqlens_k),
            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
        )


# Copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Qwen2
class Qwen2SdpaAttention(Qwen2Attention):
    """
    Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from Qwen2Attention.forward
    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "Qwen2Model is using Qwen2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)

        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
        # Reference: https://github.com/pytorch/pytorch/issues/112577.
        if query_states.device.type == "cuda" and attention_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(bsz, q_len, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


QWEN2_ATTENTION_CLASSES = {
    "eager": Qwen2Attention,
    "flash_attention_2": Qwen2FlashAttention2,
    "sdpa": Qwen2SdpaAttention,
}


class Qwen2DecoderLayer(nn.Module):
    def __init__(self, config: Qwen2Config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        self.mlp = Qwen2MLP(config)
        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Tuple[torch.Tensor]] = None,
            output_attentions: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. "
                "Please make sure use `attention_mask` instead.`"
            )
        """
        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, sequence_length)` where padding elements are indicated by 0.
            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

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


QWEN2_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 ([`Qwen2Config`]):
            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 Qwen2 Model outputting raw hidden-states without any specific head on top.",
    QWEN2_START_DOCSTRING,
)
class Qwen2PreTrainedModel(PreTrainedModel):
    config_class = Qwen2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen2DecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

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


QWEN2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

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

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

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

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

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

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

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

            Two formats are allowed:
            - a [`~cache_utils.Cache`] instance;
            - 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)`). This is also known as the legacy
            cache format.

            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
            legacy cache format will be returned.

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


@add_start_docstrings(
    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
    QWEN2_START_DOCSTRING,
)
class Qwen2Model(Qwen2PreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]

    Args:
        config: Qwen2Config
    """

    def __init__(self, config: Qwen2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.gradient_checkpointing = False
        # 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

    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
    def forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            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

        # 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:
            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 decoder_input_ids or decoder_inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        past_key_values_length = 0

        if use_cache:
            use_legacy_cache = not isinstance(past_key_values, Cache)
            if use_legacy_cache:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_usable_length(seq_length)

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

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

        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
            if is_padding_right:
                raise ValueError(
                    "You are attempting to perform batched generation with padding_side='right'"
                    " this may lead to unexpected behaviour for Flash Attention version of Qwen2. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )

        if self._attn_implementation == "flash_attention_2":
            # 2d mask is passed through the layers
            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        elif self._attn_implementation == "sdpa" and not output_attentions:
            # output_attentions=True can not be supported when using SDPA, and we fall back on
            # the manual implementation that requires a 4D causal mask in all cases.
            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
                attention_mask,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
            )
        else:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_causal_attention_mask(
                attention_mask,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
                sliding_window=self.config.sliding_window,
            )

        hidden_states = inputs_embeds

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

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

            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],)

        hidden_states = self.norm(hidden_states)

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

        next_cache = None
        if use_cache:
            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache

        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,
        )


class Qwen2ForCausalLM(Qwen2PreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = Qwen2Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

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

    def set_input_embeddings(self, value):
        self.model.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

    def get_decoder(self):
        return self.model

    def calc_loss(self, hidden_states, labels, use_cache):
        logits = self.lm_head(hidden_states)
        logits = logits.float()

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            # loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            # loss = loss_fct(shift_logits, shift_labels)
            loss = fast_cross_entropy_loss(shift_logits, shift_labels)

        if use_cache:
            return loss, logits
        else:
            return loss, None

    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
    @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,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        r"""
        Args:
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, Qwen2ForCausalLM

        >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)

        >>> prompt = "Hey, are you conscious? 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 conscious? Can you talk to me?\nI'm not conscious, 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(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]

        loss, logits = self.calc_loss(hidden_states, labels, use_cache)

        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_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        during_generate = input_ids.shape[1] > 0
        if during_generate:
            attention_mask = torch.ones(1, past_key_values[0][0].shape[2] + 1, dtype=torch.bool, device=attention_mask.device)
        elif past_key_values: # this case, attention_mask is decided by inputs_embeds
            attention_mask = torch.ones(1, past_key_values[0][0].shape[2] + inputs_embeds.shape[1], dtype=torch.bool, device=attention_mask.device)

        # Omit tokens covered by past_key_values
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                cache_length = past_key_values.get_seq_length()
                past_length = past_key_values.seen_tokens
                max_cache_length = past_key_values.get_max_length()
            else:
                cache_length = past_length = past_key_values[0][0].shape[2]
                max_cache_length = None

            # Keep only the unprocessed tokens:
            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
            # input)
            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length):]
            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
            # input_ids based on the past_length.
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.

            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
            if (
                    max_cache_length is not None
                    and attention_mask is not None
                    and cache_length + input_ids.shape[1] > max_cache_length
            ):
                attention_mask = attention_mask[:, -max_cache_length:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, past_key_values[0][0].shape[2]:]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if input_ids is not None and input_ids.shape[1] != 0:
            model_inputs = {"input_ids": input_ids}
        else:
            model_inputs = {"inputs_embeds": inputs_embeds}
        
        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        return model_inputs

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





class LlavaQwenModel(LlavaMetaModel, Qwen2Model):
    config_class = LlavaQwenConfig

    def __init__(self, config: Qwen2Config):
        super(LlavaQwenModel, self).__init__(config)


class LlavaQwenForCausalLM(Qwen2ForCausalLM, LlavaMetaForCausalLM):
    config_class = LlavaQwenConfig

    def __init__(self, config):
        # super(Qwen2ForCausalLM, self).__init__(config)
        Qwen2ForCausalLM.__init__(self, config)
        config.model_type = "llava_qwen"
        config.rope_scaling = None

        self.model = LlavaQwenModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        # Initialize weights and apply final processing
        self.post_init()

    def get_model(self):
        return self.model

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        images: Optional[torch.FloatTensor] = None,
        image_sizes: Optional[List[List[int]]] = None,
        return_dict: Optional[bool] = None,
        modalities: Optional[List[str]] = ["image"],
        clip_sizes: Optional[List[int]] = None,
        image_sizes_per_clip: Optional[List] = None,
        dpo_forward: Optional[bool] = False,
        cache_position=None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        if inputs_embeds is None:
            (input_ids, position_ids, attention_mask, past_key_values, inputs_embeds, labels) = (
                self.prepare_inputs_labels_for_multimodal_interleave(
                    input_ids=input_ids,
                    position_ids=position_ids,
                    attention_mask=attention_mask,
                    past_key_values=past_key_values,
                    labels=labels,
                    images=images,
                    modalities=modalities,
                    clip_sizes=clip_sizes,
                    image_sizes_per_clip=image_sizes_per_clip,
                )
            )
        if dpo_forward:
            outputs = self.model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                inputs_embeds=inputs_embeds,
                use_cache=use_cache,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )

            hidden_states = outputs[0]
            logits = self.lm_head(hidden_states)
            return logits, labels

        else:
            return super().forward(
                input_ids=input_ids,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                inputs_embeds=inputs_embeds,
                labels=labels,
                use_cache=use_cache,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )

    @torch.no_grad()
    def generate(
        self,
        inputs: Optional[torch.Tensor] = None,
        images: Optional[torch.Tensor] = None,
        image_sizes: Optional[torch.Tensor] = None,
        modalities: Optional[List[str]] = ["image"],
        clip_sizes: Optional[List] = None,
        image_sizes_per_clip: Optional[List] = None,
        **kwargs,
    ) -> Union[GenerateOutput, torch.LongTensor]:
        position_ids = kwargs.pop("position_ids", None)
        attention_mask = kwargs.pop("attention_mask", None)
        if "inputs_embeds" in kwargs:
            raise NotImplementedError("`inputs_embeds` is not supported")

        if images is not None:
            (inputs, position_ids, attention_mask, _, inputs_embeds, _) = (
                self.prepare_inputs_labels_for_multimodal_interleave(
                    input_ids=inputs,
                    position_ids=position_ids,
                    attention_mask=attention_mask,
                    past_key_values=None,
                    labels=None,
                    images=images,
                    modalities=modalities,
                    clip_sizes=clip_sizes,
                    image_sizes_per_clip=image_sizes_per_clip,
                )
            )
        else:
            inputs_embeds = self.get_model().embed_tokens(inputs)

        return super().generate(position_ids=position_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, **kwargs)

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
        images = kwargs.pop("images", None)
        image_sizes = kwargs.pop("image_sizes", None)
        inputs = super().prepare_inputs_for_generation(
            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
        )
        if images is not None:
            inputs["images"] = images
        if image_sizes is not None:
            inputs["image_sizes"] = image_sizes
        return inputs