models/modeling_qwen2_vl_fast.py

import os
import math
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import LayerNorm

from transformers.modeling_utils import PreTrainedModel
from transformers.configuration_utils import PretrainedConfig
from transformers.modeling_rope_utils import rope_config_validation, ROPE_INIT_FUNCTIONS
from transformers.cache_utils import Cache, SlidingWindowCache, StaticCache
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
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 transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    ModelOutput,
)
from transformers.activations import ACT2FN
from transformers.generation import GenerationMixin

if is_flash_attn_2_available():
    from flash_attn import flash_attn_varlen_func

    from transformers.modeling_flash_attention_utils import _flash_attention_forward
else:
    flash_attn_varlen_func = None

# from apex.normalization.fused_layer_norm import fused_rms_norm_affine

logger = logging.get_logger(__name__)

@dataclass
class Qwen2VLCausalLMOutputWithPast(ModelOutput):
    """
    Base class for Qwen2VL causal language model (or autoregressive) outputs.

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

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

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

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
    """

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

class Qwen2VLVisionConfig(PretrainedConfig):
    model_type = "qwen2_vl"

    def __init__(
        self,
        depth=32,
        embed_dim=1280,
        hidden_size=3584,
        hidden_act="quick_gelu",
        mlp_ratio=4,
        num_heads=16,
        in_channels=3,
        patch_size=14,
        spatial_merge_size=2,
        temporal_patch_size=2,
        attn_implementation='flash_attention_2',
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.depth = depth
        self.embed_dim = embed_dim
        self.hidden_size = hidden_size
        self.hidden_act = hidden_act
        self.mlp_ratio = mlp_ratio
        self.num_heads = num_heads
        self.in_channels = in_channels
        self.patch_size = patch_size
        self.spatial_merge_size = spatial_merge_size
        self.temporal_patch_size = temporal_patch_size
        self.attn_implementation = attn_implementation

    @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)

        if config_dict.get("model_type") == "qwen2_vl":
            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:
            logger.warning(
                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)


class Qwen2VLConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2VLModel`]. It is used to instantiate a
    Qwen2-VL model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of
    Qwen2-VL-7B-Instruct [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct).

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 152064):
            Vocabulary size of the Qwen2VL model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`Qwen2VLModel`]
        hidden_size (`int`, *optional*, defaults to 8192):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 29568):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 80):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 64):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 8):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 32768):
            The maximum sequence length that this model might ever be used with.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            The base period of the RoPE embeddings.
        use_sliding_window (`bool`, *optional*, defaults to `False`):
            Whether to use sliding window attention.
        sliding_window (`int`, *optional*, defaults to 4096):
            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
        max_window_layers (`int`, *optional*, defaults to 80):
            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        vision_config (`Dict`, *optional*):
            The config for the visual encoder initialization.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
            accordingly.
            Expected contents:
                `rope_type` (`str`):
                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
                    'llama3'], with 'default' being the original RoPE implementation.
                `factor` (`float`, *optional*):
                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
                    original maximum pre-trained length.
                `original_max_position_embeddings` (`int`, *optional*):
                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
                    pretraining.
                `attention_factor` (`float`, *optional*):
                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
                    computation. If unspecified, it defaults to value recommended by the implementation, using the
                    `factor` field to infer the suggested value.
                `beta_fast` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
                    ramp function. If unspecified, it defaults to 32.
                `beta_slow` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
                    ramp function. If unspecified, it defaults to 1.
                `short_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `long_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `low_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
                `high_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE

    ```python
    >>> from transformers import Qwen2VLForConditionalGeneration, Qwen2VLConfig

    >>> # Initializing a Qwen2VL style configuration
    >>> configuration = Qwen2VLConfig()

    >>> # Initializing a model from the Qwen2-VL-7B style configuration
    >>> model = Qwen2VLForConditionalGeneration(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen2_vl"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=152064,
        hidden_size=8192,
        intermediate_size=29568,
        num_hidden_layers=80,
        num_attention_heads=64,
        num_key_value_heads=8,
        hidden_act="silu",
        max_position_embeddings=32768,
        initializer_range=0.02,
        rms_norm_eps=1e-05,
        use_cache=True,
        tie_word_embeddings=False,
        rope_theta=1000000.0,
        use_sliding_window=False,
        sliding_window=4096,
        max_window_layers=80,
        attention_dropout=0.0,
        rope_scaling=None,
        spatial_merge_size=2,
        attn_implementation='flash_attention_2',
        **kwargs,
    ):

        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        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.use_sliding_window = use_sliding_window
        self.sliding_window = sliding_window
        self.max_window_layers = max_window_layers

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.rope_scaling = rope_scaling
        self.spatial_merge_size = spatial_merge_size
        self.attn_implementation = attn_implementation

        # Validate the correctness of rotary position embeddings parameters
        # BC: if there is a 'type' field, move it to 'rope_type'.
        # and change type from 'mrope' to 'default' because `mrope` does defeault RoPE calculations
        # one can set it to "linear"/"dynamic" etc. to have scaled RoPE
        # TODO: @raushan update config in the hub
        if self.rope_scaling is not None and "type" in self.rope_scaling:
            if self.rope_scaling["type"] == "mrope":
                self.rope_scaling["type"] = "default"
            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
        rope_config_validation(self, ignore_keys={"mrope_section"})

        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

# 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)

def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).

    Explanation:
        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
        vision embedding part, we apply rotary position embedding on temporal, height and width dimension seperately.
        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
        difference with modern LLMs.

    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.
        mrope_section(`List(int)`):
            Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
        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.
    """
    mrope_section = mrope_section * 2
    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )
    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )

    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

def apply_rotary_pos_emb_vision(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
    orig_dtype = tensor.dtype
    tensor = tensor.float()
    cos = freqs.cos()
    sin = freqs.sin()
    cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
    sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
    output = (tensor * cos) + (rotate_half(tensor) * sin)
    output = output.to(orig_dtype)
    return output


class VisionRotaryEmbedding(nn.Module):
    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs
    
class PatchEmbed(nn.Module):
    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 2,
        in_channels: int = 3,
        embed_dim: int = 1152,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.in_channels = in_channels
        self.embed_dim = embed_dim

        kernel_size = [temporal_patch_size, patch_size, patch_size]
        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


class PatchMerger(nn.Module):
    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        self.ln_q = LayerNorm(context_dim, eps=1e-6)
        self.mlp = nn.Sequential(
            nn.Linear(self.hidden_size, self.hidden_size),
            nn.GELU(),
            nn.Linear(self.hidden_size, dim),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
        return x
    
class VisionMlp(nn.Module):
    def __init__(self, dim: int, hidden_dim: int, hidden_act: str) -> None:
        super().__init__()
        self.fc1 = nn.Linear(dim, hidden_dim)
        self.act = ACT2FN[hidden_act]
        self.fc2 = nn.Linear(hidden_dim, dim)

    def forward(self, x) -> torch.Tensor:
        return self.fc2(self.act(self.fc1(x)))


class VisionAttention(nn.Module):
    def __init__(self, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.qkv = nn.Linear(dim, dim * 3, bias=True)
        self.proj = nn.Linear(dim, dim)

    def forward(
        self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
        k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)

        attention_mask = torch.full(
            [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
        )
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = 0

        q = q.transpose(0, 1)
        k = k.transpose(0, 1)
        v = v.transpose(0, 1)
        attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
        attn_weights = attn_weights + attention_mask
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.transpose(0, 1)
        attn_output = attn_output.reshape(seq_length, -1)
        attn_output = self.proj(attn_output)
        return attn_output


class VisionFlashAttention2(nn.Module):
    def __init__(self, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.qkv = nn.Linear(dim, dim * 3, bias=True)
        self.proj = nn.Linear(dim, dim)

    def forward(
        self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
        k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)

        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
        attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
            seq_length, -1
        )
        attn_output = self.proj(attn_output)
        return attn_output

QWEN2_VL_VISION_ATTENTION_CLASSES = {
    "eager": VisionAttention,
    "flash_attention_2": VisionFlashAttention2,
}


class Qwen2VLVisionBlock(nn.Module):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = LayerNorm(config.embed_dim, eps=1e-6)
        self.norm2 = LayerNorm(config.embed_dim, eps=1e-6)
        mlp_hidden_dim = int(config.embed_dim * config.mlp_ratio)

        self.attn = QWEN2_VL_VISION_ATTENTION_CLASSES[attn_implementation](
            config.embed_dim, num_heads=config.num_heads
        )
        self.mlp = VisionMlp(dim=config.embed_dim, hidden_dim=mlp_hidden_dim, hidden_act=config.hidden_act)

    def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states), cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states

class Qwen2VLPreTrainedModel(PreTrainedModel):
    config_class = Qwen2VLConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen2VLDecoderLayer", "Qwen2VLVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = False
    _supports_cache_class = True
    _supports_static_cache = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, (nn.Linear, nn.Conv3d)):
            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_()
    
class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
    config_class = Qwen2VLVisionConfig
    _no_split_modules = ["Qwen2VLVisionBlock"]

    def __init__(self, config) -> None:
        super().__init__(config)
        self.spatial_merge_size = config.spatial_merge_size

        self.patch_embed = PatchEmbed(
            patch_size=config.patch_size,
            temporal_patch_size=config.temporal_patch_size,
            in_channels=config.in_channels,
            embed_dim=config.embed_dim,
        )

        head_dim = config.embed_dim // config.num_heads
        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList(
            [Qwen2VLVisionBlock(config, config.attn_implementation) for _ in range(config.depth)]
        )
        self.merger = PatchMerger(
            dim=config.hidden_size, context_dim=config.embed_dim, spatial_merge_size=config.spatial_merge_size
        )
        # Initialize weights and apply final processing
        self.gradient_checkpointing = False
        self.post_init()

    def get_dtype(self) -> torch.dtype:
        return self.blocks[0].mlp.fc2.weight.dtype

    def get_device(self) -> torch.device:
        return self.blocks[0].mlp.fc2.weight.device

    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
        hidden_states = self.patch_embed(hidden_states)
        rotary_pos_emb = self.rot_pos_emb(grid_thw)

        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0, dtype=torch.int32
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

        for blk in self.blocks:
            if self.gradient_checkpointing and self.training:
                hidden_states = self._gradient_checkpointing_func(
                    blk.__call__,
                    hidden_states,
                    cu_seqlens,
                    rotary_pos_emb,
                )
            else:
                hidden_states = blk(hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb)

        return self.merger(hidden_states)

# 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
#         self.normalized_shape = torch.Size((hidden_size, ))

#     def forward(self, hidden_states):
#         return fused_rms_norm_affine(input=hidden_states,
#                                      weight=self.weight,
#                                      normalized_shape=self.normalized_shape,
#                                      eps=self.variance_epsilon,
#                                      memory_efficient=True)

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):
        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)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
    
class Qwen2VLRotaryEmbedding(nn.Module):
    def __init__(
        self,
        dim=None,
        max_position_embeddings=2048,
        base=10000,
        device=None,
        scaling_factor=1.0,
        rope_type="default",
        config: Optional[Qwen2VLConfig] = None,
    ):
        super().__init__()
        # TODO (joao): remove the `if` below, only used for BC
        self.rope_kwargs = {}
        if config is None:
            logger.warning_once(
                "`Qwen2VLRotaryEmbedding` can now be fully parameterized by passing the model config through the "
                "`config` argument. All other arguments will be removed in v4.46"
            )
            self.rope_kwargs = {
                "rope_type": rope_type,
                "factor": scaling_factor,
                "dim": dim,
                "base": base,
                "max_position_embeddings": max_position_embeddings,
            }
            self.rope_type = rope_type
            self.max_seq_len_cached = max_position_embeddings
            self.original_max_seq_len = max_position_embeddings
        else:
            # BC: "rope_type" was originally "type"
            if config.rope_scaling is not None:
                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
            else:
                self.rope_type = "default"
            self.max_seq_len_cached = config.max_position_embeddings
            self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(
                self.config, device, seq_len=seq_len, **self.rope_kwargs
            )
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        position_ids = position_ids.permute(2, 0, 1)
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block. In contrast to other models, Qwen2_VL has different position ids for thw grids
        # So we expand the inv_freq to shape (3, ...)
        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

# Copied from transformers.models.qwen2.modeling_qwen2.Qwen2MLP
class Qwen2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        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, hidden_state):
        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))

# 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 Qwen2VLAttention(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: Qwen2VLConfig, 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
        self.rope_scaling = config.rope_scaling

        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)


class Qwen2VLFlashAttention2(Qwen2VLAttention):
    """
    Qwen2VL flash attention module, following Qwen2VL attention module. This module inherits from `Qwen2VLAttention`
    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.
    """

    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,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
        use_rmpad: Optional[bool] = False,
        cu_seqlens: Optional[torch.Tensor] = False,
    ):
        """
        Train: 
          unpad: (bsz, q_len) = (1, acc_seqlen)
          pad: (bsz, q_len) = (bsz, q_len)
        Test:
          first_iter: (bsz, q_len) = (bsz, q_len)
          other: (bsz, q_len) = (bsz, 1)
        """
        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)

        cos, sin = position_embeddings

        query_states, key_states = apply_multimodal_rotary_pos_emb(
            query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
        )

        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)
        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)

        if use_rmpad:
            max_seqlen = torch.max(cu_seqlens[1:] - cu_seqlens[:-1]).item()
            attn_output = flash_attn_varlen_func(
                query_states.squeeze(0), key_states.squeeze(0), value_states.squeeze(0), 
                cu_seqlens_q=cu_seqlens, cu_seqlens_k=cu_seqlens, 
                max_seqlen_q=max_seqlen, max_seqlen_k=max_seqlen,
                dropout_p=dropout_rate,
                causal=self.is_causal, window_size=(-1, -1),
            )
        else:
            attn_output = _flash_attention_forward(
                query_states, key_states, value_states,
                attention_mask,
                q_len,
                dropout=dropout_rate,
                sliding_window=None,
                is_causal=self.is_causal,
                use_top_left_mask=self._flash_attn_uses_top_left_mask,
            )

        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

QWEN2_VL_ATTENTION_CLASSES = {
    "flash_attention_2": Qwen2VLFlashAttention2,
}

class Qwen2VLDecoderLayer(nn.Module):
    def __init__(self, config: Qwen2VLConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.attn_implementation != "flash_attention_2":
            logger.error(
                f"只支持 flash_attention_2！config.attn_implementation={config.attn_implementation}"
            )
        self.self_attn = QWEN2_VL_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,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
        use_rmpad: Optional[bool] = False,
        cu_seqlens: Optional[torch.Tensor] = False,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, 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
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """

        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,
            position_embeddings=position_embeddings,
            use_rmpad=use_rmpad,
            cu_seqlens=cu_seqlens,
        )
        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

class Qwen2VLModel(Qwen2VLPreTrainedModel):
    def __init__(self, config: Qwen2VLConfig):
        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([Qwen2VLDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Qwen2VLRotaryEmbedding(config=config)

        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

    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,
        use_rmpad: Optional[bool] = False,
        cu_seqlens: Optional[torch.Tensor] = False,
    ) -> 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

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or 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
        

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # 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,
                    position_embeddings,
                    use_rmpad,
                    cu_seqlens,
                )
            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,
                    position_embeddings=position_embeddings,
                    use_rmpad=use_rmpad,
                    cu_seqlens=cu_seqlens,
                )

            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 = next_decoder_cache if use_cache else None

        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )
    
class Qwen2VLForCausalLM(Qwen2VLPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = Qwen2VLModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.padding_side = "left"  # set it to left by default, user can use setter to change padding_sides

        # 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 get_rope_index(
        self,
        input_ids: torch.LongTensor,
        image_grid_thw: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Calculate the 3D rope index based on image and video's temporal, height and width in LLM.

        Explanation:
            Each embedding sequence contains vision embedding and text embedding or just contains text embedding.

            For pure text embedding sequence, the rotary position embedding has no difference with mordern LLMs.
            Examples:
                input_ids: [T T T T T], here T is for text.
                temporal position_ids: [0, 1, 2, 3, 4]
                height position_ids: [0, 1, 2, 3, 4]
                width position_ids: [0, 1, 2, 3, 4]

            For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
            and 1D rotary position embeddin for text part.
            Examples:
                Assume we have a video input with 3 temporal patches, 2 height patches and 2 width patches.
                input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
                vision temporal position_ids: [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2]
                vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
                vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
                text temporal position_ids: [3, 4, 5, 6, 7]
                text height position_ids: [3, 4, 5, 6, 7]
                text width position_ids: [3, 4, 5, 6, 7]
                Here we calculate the text start position_ids as the max vision position_ids plus 1.

        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.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
            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**.

        Returns:
            position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
            mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
        """
        spatial_merge_size = self.config.spatial_merge_size
        vision_token_id = self.config.image_token_id
        vision_start_token_id = self.config.vision_start_token_id
        assert image_grid_thw is not None # TODO：测试纯文本会不会卡住
        total_input_ids = input_ids
        position_ids = torch.ones(
            3, input_ids.shape[0], input_ids.shape[1], dtype=input_ids.dtype, device=input_ids.device
        )
        vision_index = 0
        for i, input_ids in enumerate(total_input_ids):
            if attention_mask is not None:
                input_ids = input_ids[attention_mask[i] == 1]
            vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
            vision_num = (input_ids[vision_start_indices + 1] == vision_token_id).sum()
            input_tokens = input_ids.tolist()
            llm_pos_ids_list: list = []
            st = 0
            remain_vision_num = vision_num
            for _ in range(vision_num):
                if vision_token_id in input_tokens and remain_vision_num > 0:
                    ed_vision = input_tokens.index(vision_token_id, st)
                else:
                    ed_vision = len(input_tokens) + 1

                t, h, w = (
                    image_grid_thw[vision_index][0],
                    image_grid_thw[vision_index][1],
                    image_grid_thw[vision_index][2],
                )
                vision_index += 1
                remain_vision_num -= 1
                ed = ed_vision
                
                llm_grid_t, llm_grid_h, llm_grid_w = (
                    t.item(),
                    h.item() // spatial_merge_size,
                    w.item() // spatial_merge_size,
                )
                text_len = ed - st

                st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

                t_index = torch.arange(llm_grid_t).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten()
                h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
                w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
                llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
                st = ed + llm_grid_t * llm_grid_h * llm_grid_w

            if st < len(input_tokens):
                st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                text_len = len(input_tokens) - st
                llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

            llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
            position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
        position_ids = position_ids.permute(1, 2, 0)
        return position_ids

    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,
        use_rmpad: Optional[bool] = False,
        cu_seqlens: Optional[torch.Tensor] = False,
    ) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:

        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


        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,
            use_rmpad=use_rmpad,
            cu_seqlens=cu_seqlens,
        )

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

        if not return_dict:
            output = (logits,) + outputs[1:]
            return output

        return Qwen2VLCausalLMOutputWithPast(
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )