modeling_hyper_qwen2.py

# coding=utf-8
# 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
from typing import List, Optional, Tuple, Union

from einops import rearrange, repeat
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
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, SequenceClassifierOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.models.qwen2.modeling_qwen2 import Qwen2Attention, Qwen2FlashAttention2, Qwen2SdpaAttention
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 .configuration_hyper_qwen2 import HyperQwen2Config




try:
    from flash_attn.layers.rotary import apply_rotary_emb_func
    from einops import rearrange

    use_flash_rotary = True
    print("use flash_attn rotary")
except ImportError:
    use_flash_rotary = False
    print("import flash_attn rotary fail")


try:
    from torch.nn.attention.flex_attention import create_block_mask
    from torch.nn.attention.flex_attention import flex_attention
    flex_attention = torch.compile(flex_attention, dynamic=False)
except ImportError:
    pass

logger = logging.get_logger(__name__)


_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta"
_CONFIG_FOR_DOC = "HyperQwen2Config"


# 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):
        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),
        )

class RotaryEmbedding(torch.nn.Module):
    def __init__(self, dim, base=10000, use_fp32=False, use_outer_in_rope=False):
        super().__init__()
        self.dim = dim
        self.base = base
        self.use_fp32 = use_fp32
        if use_fp32:
            self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        else:
            inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
            self.register_buffer("inv_freq", inv_freq)
      
        self._rotary_pos_emb_cache = None
        self._seq_len_cached = 0
        self.use_outer_in_rope = use_outer_in_rope
        self._ntk_alpha_cached = 1.0

    def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
        seqlen = max_seq_len + offset
        if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
            base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
            self.inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, device=self.inv_freq.device).float() / self.dim))
            self._seq_len_cached = seqlen
            self._ntk_alpha_cached = ntk_alpha
            seq = torch.arange(seqlen, device=self.inv_freq.device)
            # Don't do einsum, it converts fp32 to fp16            # TODO: CHECK this
            if self.use_outer_in_rope:
                freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
            else:
                freqs = einsum('i , j -> i j', seq.type_as(self.inv_freq), self.inv_freq)
            # first part even vector components, second part odd vector components,
            #  2 * dim in dimension size
            emb = torch.cat((freqs, freqs), dim=-1)
            # emb [seq_length, .., dim]
            from einops import rearrange
            self._rotary_pos_emb_cache = rearrange(emb, 'n d -> n 1 1 d')

    def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
        self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
        return self._rotary_pos_emb_cache[offset:offset + max_seq_len]

# 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)




def _rotate_half(x):
    """
    change sign so the last dimension becomes [-odd, +even]
    """
    from einops import rearrange
    x = rearrange(x, '... (j d) -> ... j d', j=2)
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1)

def apply_rotary_pos_emb_core(t, freqs, use_fp32=False, debug=False):
    """
    input tensor t is of shape [seq_length, ..., dim]
    rotary positional embeding tensor freqs is of shape [seq_length, ..., dim]
    check https://kexue.fm/archives/8265 for detailed formulas
    """

    if use_flash_rotary and use_fp32:
        t_ = rearrange(t, 's b ... -> b s ...').contiguous()
        if use_fp32:
            t_ = t_.float()
        freqs = freqs.squeeze(1).squeeze(1)
        cos = freqs[:, :freqs.shape[-1] // 2].cos()
        sin = freqs[:, :freqs.shape[-1] // 2].sin()
        output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
        return rearrange(output, 'b s ... -> s b ...')

    rot_dim = freqs.shape[-1]
    # ideally t_pass is empty so rotary pos embedding is applied to all tensor t
    t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
    
    if use_fp32:
        t_ = t_.float()
        t_pass_ = t_pass_.float()
    # first part is cosine component
    # second part is sine component, need to change signs with _rotate_half method
    t_ = (t_ * freqs.cos()) + (_rotate_half(t_) * freqs.sin())
    return torch.cat((t_, t_pass_), dim=-1).type_as(t)

class HyperQwen2Attention(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: HyperQwen2Config, layer_idx: Optional[int] = None, is_hyper_enabled=False):
        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,
        )
        self.rotary_emb_core = RotaryEmbedding(
            self.head_dim, base=self.rope_theta, use_fp32=True, use_outer_in_rope=True
        )
        # Hyper Attention Modules
        self.is_hyper_enabled = is_hyper_enabled
        if self.is_hyper_enabled:
            self.v_kv_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim * 2, bias=True)
          
            self.visual_cache={}
        
        self.use_flexattention = True



    def apply_mi_rope(self, key_layer, image_pos, length_each_img):
        # input shape should be [s b h d]
        key_layer = rearrange(key_layer, 'b h s d -> s b h d')
        if self.rotary_emb_core.inv_freq.device!=key_layer.device:
            self.rotary_emb_core.inv_freq = self.rotary_emb_core.inv_freq.to(key_layer.device)
        rotary_pos_emb_max_seq_len = self.config.max_position_embeddings
        ntk_alpha = 1
        rotary_pos_emb = self.rotary_emb_core(rotary_pos_emb_max_seq_len, ntk_alpha=ntk_alpha)
        assert rotary_pos_emb is not None

        if isinstance(rotary_pos_emb, tuple):
            rotary_pos_emb = rotary_pos_emb
        else:
            rotary_pos_emb = ((rotary_pos_emb,) * 2)
 

        if rotary_pos_emb is not None:
            q_pos_emb, k_pos_emb = rotary_pos_emb
            # ic(key_layer.shape, k_pos_emb.shape)
            
   
            k_pos_emb = repeat(k_pos_emb[image_pos], 'N_img b h d -> (N_img L) b h d', L=length_each_img) # N_img, dim

            key_layer = apply_rotary_pos_emb_core(key_layer, k_pos_emb, use_fp32=True) # TODO difference
        key_layer = rearrange(key_layer, 's b h d -> b h s d')
        return key_layer


# def hyper_mask_always_true(b, h, q_idx, kv_idx):
#     return q_idx>=0

# def causal(b, h, q_idx, kv_idx):
#     return q_idx >= kv_idx


# def create_hyper_attention(media_starts_extend, q_len, kv_len, each_visual_len):
    
#     visual_len = kv_len - q_len
#     def hyper_mask_dynamic(b, h, q_idx, kv_idx):
#         return torch.where(kv_idx<visual_len, q_idx>=media_starts_extend[kv_idx], causal(b, h, q_idx, kv_idx-visual_len))
    
#     return create_block_mask(hyper_mask_dynamic, B=None, H=None, Q_LEN=q_len, KV_LEN=kv_len, BLOCK_SIZE=128, _compile=True)

# Copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Qwen2
class HyperQwen2SdpaAttention(HyperQwen2Attention):
    """
    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.
    """

    def hyperattention(self,hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        image_embeds=None,
        media_offset=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]]]:
        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)

        # add visual to kv
        length_each_img = image_embeds.shape[1]
        image_embeds = self.v_kv_proj(image_embeds)
        image_start = 0
        context_layer = []                 
        for bi, media_starts in enumerate(media_offset):
            num_images = media_starts.shape[0]
            if num_images > 0:
                if q_len == 1:
                    full_mask = torch.ones((1,1,1, num_images*length_each_img + kv_seq_len)).bool().to(query_states.device)
                else:
                    causal_mask = torch.tril(torch.ones(q_len, kv_seq_len, dtype=torch.bool, device=query_states.device)).bool()
                    # 扩展维度以匹配 (bsz, 1, q_len, kv_seq_len)
                    causal_mask = causal_mask.unsqueeze(0).unsqueeze(0)

                    matrix = torch.arange(q_len, device=media_offset[0].device).reshape(-1,1)
                    t2vmask = ~(matrix<media_starts.view(1, -1)) 
                    t2vmask = repeat(t2vmask, 'seq_t seq_v -> 1 1 seq_t (seq_v v_token)', v_token=length_each_img).to(query_states.device)
                    full_mask = torch.cat([t2vmask, causal_mask], dim=3) # unsqueeze batch dim (batch, 1, seq_q, seq_k)

                curr_query_layer = query_states[bi:bi+1]
                # order is sbhd
                curr_visual_key_layer, curr_visual_value_layer = rearrange(image_embeds[image_start:image_start+num_images], 'BL Lv (H KV D) -> KV 1 H (BL Lv) D', KV=2, H=self.num_key_value_heads) # b h s d
                image_start += num_images
                # ic(media_starts)
                curr_visual_key_layer = self.apply_mi_rope(curr_visual_key_layer, media_starts, length_each_img=length_each_img)

                curr_visual_key_layer = repeat_kv(curr_visual_key_layer, self.num_key_value_groups)
                curr_visual_value_layer = repeat_kv(curr_visual_value_layer, self.num_key_value_groups)

                curr_key_layer = torch.cat([curr_visual_key_layer, key_states[bi:bi+1]], dim=2)
                curr_value_layer = torch.cat([curr_visual_value_layer, value_states[bi:bi+1]], dim=2)
                is_causal = False
            else:
                # 执行无图attention
                curr_query_layer = query_states[bi:bi+1]
                curr_key_layer = key_states[bi:bi+1]
                curr_value_layer = value_states[bi:bi+1]
                full_mask = causal_mask
                is_causal = True if causal_mask is None and q_len > 1 else False
                if is_causal:
                    full_mask = None
                else:
                    causal_mask = torch.tril(torch.ones(q_len, kv_seq_len, dtype=torch.bool, device=query_states.device)).bool()
                    # 扩展维度以匹配 (bsz, 1, q_len, kv_seq_len)
                    causal_mask = causal_mask.unsqueeze(0).unsqueeze(0)
                    full_mask = causal_mask


            # 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 curr_query_layer.device.type == "cuda" and full_mask is not None:
                curr_query_layer = curr_query_layer.contiguous()
                curr_key_layer = curr_key_layer.contiguous()
                curr_value_layer = curr_value_layer.contiguous()
           
            attn_output = torch.nn.functional.scaled_dot_product_attention(
                curr_query_layer, # (batch, ..., sequence, dim)
                curr_key_layer,
                curr_value_layer,
                attn_mask=full_mask, # (N, ..., L, S) A boolean mask where a value of True indicates that the element *should* take part in attention.
                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=is_causal,
                # enable_gqa=True, # gqa can not be used because mask requires XFORMERS and not support gqa
            ) # -> (N, ..., L, Ev)
            assert attn_output.shape[0] == 1
            context_layer.append(attn_output)
        attn_output = context_layer = torch.cat(context_layer, dim=0)

        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

    # def hyperattention_flex(self,hidden_states: torch.Tensor,
    #     attention_mask: Optional[torch.Tensor] = None,
    #     position_ids: Optional[torch.LongTensor] = None,
    #     image_embeds=None,
    #     media_offset=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]]]:
    #     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)

    #     # add visual to kv
    #     length_each_img = image_embeds.shape[1]
    #     image_embeds = self.v_kv_proj(image_embeds)
    #     image_start = 0
    #     context_layer = []           

        
    #     for bi, media_starts in enumerate(media_offset):
    #         num_images = media_starts.shape[0]
    #         if num_images > 0:
    #             if q_len == 1:
    #                 hyper_maks = create_block_mask(hyper_mask_always_true, B=None, H=None, Q_LEN=q_len, KV_LEN=kv_seq_len+len(media_starts)*length_each_img)
    #             else:
    #                 media_starts_extend = repeat(media_starts, 'seq_v -> (seq_v v_token)', v_token=length_each_img)
    #                 extend_len = media_starts_extend.shape[0]+kv_seq_len
    #                 if extend_len%128!=0:
    #                     extend_len = (extend_len//128+1)*128
    #                 extend_len = extend_len-media_starts_extend.shape[0]
    #                 media_starts_extend = torch.cat([media_starts_extend, torch.zeros(extend_len, device=media_starts_extend.device, dtype=media_starts_extend.dtype)],dim=0)
    #                 hyper_maks = create_hyper_attention(media_starts_extend, q_len, kv_seq_len+len(media_starts)*length_each_img, length_each_img)

    #             curr_query_layer = query_states[bi:bi+1]
    #             # order is sbhd
    #             curr_visual_key_layer, curr_visual_value_layer = rearrange(image_embeds[image_start:image_start+num_images], 'BL Lv (H KV D) -> KV 1 H (BL Lv) D', KV=2, H=self.num_key_value_heads) # b h s d
    #             image_start += num_images
    #             # ic(media_starts)
    #             curr_visual_key_layer = self.apply_mi_rope(curr_visual_key_layer, media_starts, length_each_img=length_each_img)

    #             curr_visual_key_layer = repeat_kv(curr_visual_key_layer, self.num_key_value_groups)
    #             curr_visual_value_layer = repeat_kv(curr_visual_value_layer, self.num_key_value_groups)

    #             curr_key_layer = torch.cat([curr_visual_key_layer, key_states[bi:bi+1]], dim=2)
    #             curr_value_layer = torch.cat([curr_visual_value_layer, value_states[bi:bi+1]], dim=2)
    #             is_causal = False
    #         else:
    #             # 执行无图attention
    #             curr_query_layer = query_states[bi:bi+1]
    #             curr_key_layer = key_states[bi:bi+1]
    #             curr_value_layer = value_states[bi:bi+1]
    #             full_mask = causal_mask
    #             is_causal = True if causal_mask is None and q_len > 1 else False
    #             if is_causal:
    #                 hyper_maks = create_block_mask(hyper_mask_always_true, B=None, H=None, Q_LEN=q_len, KV_LEN=kv_seq_len, _compile=True)
    #             else:
    #                 hyper_maks = create_block_mask(causal, B=None, H=None, Q_LEN=q_len, KV_LEN=kv_seq_len, _compile=True)


    #         # 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 curr_query_layer.device.type == "cuda" and attention_mask is not None:
    #             curr_query_layer = curr_query_layer.contiguous()
    #             curr_key_layer = curr_key_layer.contiguous()
    #             curr_value_layer = curr_value_layer.contiguous()

            
    #         attn_output = flex_attention(
    #             curr_query_layer,
    #             curr_key_layer,
    #             curr_value_layer,
    #             block_mask=hyper_maks
    #         )
    #         # attn_output = torch.nn.functional.scaled_dot_product_attention(
    #         #     curr_query_layer, # (batch, ..., sequence, dim)
    #         #     curr_key_layer,
    #         #     curr_value_layer,
    #         #     attn_mask=full_mask, # (N, ..., L, S) A boolean mask where a value of True indicates that the element *should* take part in attention.
    #         #     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=is_causal,
    #         # ) # -> (N, ..., L, Ev)
    #         assert attn_output.shape[0] == 1
    #         context_layer.append(attn_output)
    #     attn_output = context_layer = torch.cat(context_layer, dim=0)

    #     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

    # Adapted from Qwen2Attention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        image_embeds=None,
        media_offset=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,
            )
        if self.is_hyper_enabled and image_embeds is not None:
            # return self.hyperattention_flex(hidden_states, attention_mask, position_ids, image_embeds, media_offset, past_key_value, output_attentions, use_cache)
            return self.hyperattention(hidden_states, attention_mask, position_ids, image_embeds, media_offset, past_key_value, output_attentions, 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


# Original Attention of Qwen2
QWEN2_ATTENTION_CLASSES = {
    "eager": Qwen2Attention,
    "flash_attention_2": Qwen2FlashAttention2,
    "sdpa": Qwen2SdpaAttention,
}


class HyperQwen2DecoderLayer(nn.Module):
    def __init__(self, config: HyperQwen2Config, 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.is_hyper_enabled = (layer_idx+1) in config.hyper_layers
        if self.is_hyper_enabled:
            self.self_attn = HyperQwen2SdpaAttention(config, layer_idx, is_hyper_enabled=self.is_hyper_enabled)
        else:
            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)

    @property
    def device(self):
        return self.input_layernorm.weight.device

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        image_embeds=None,
        media_offset=None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
    ) -> 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
        """
        # if hidden_states.device != self.device:
        #     hidden_states = hidden_states.to(self.device)
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        
        # Shared LayerNorm
        if image_embeds is not None and self.is_hyper_enabled:
            image_embeds = self.input_layernorm(image_embeds)
            media_kwargs = {"image_embeds": image_embeds, "media_offset": media_offset}
        else:
            image_embeds = media_offset = None
            media_kwargs = {}
        # 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,
            **media_kwargs,
        )
        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 ([`HyperQwen2Config`]):
            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 = HyperQwen2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["HyperQwen2DecoderLayer"]
    _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 HyperQwen2Model(Qwen2PreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]

    Args:
        config: HyperQwen2Config
    """

    def __init__(self, config: HyperQwen2Config):
        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(
            [HyperQwen2DecoderLayer(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,
        image_embeds=None,
        media_offset=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,
                sliding_window=self.config.sliding_window,
            )
        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

        # beam search
        if batch_size != len(media_offset):
            # The model is performing beamsearch, repeat the visual content
            beam_factor = batch_size // len(media_offset)
            assert batch_size % len(media_offset) == 0
            media_offset = media_offset * beam_factor
            image_embeds = repeat(image_embeds, 'B L D -> (factor B) L D', factor=beam_factor)

        # # Flex mask
        
        # expected_q_len =  hidden_states.shape[1]
        # expected_kv_len = expected_q_len
        # if past_key_value is not None:
        #     kv_seq_len += past_key_value.get_usable_length(kv_seq_len, 1)
        
        #     length_each_img = image_embeds.shape[1]
        #     expected_kv_len = [expected_kv_len+len(_)*length_each_img for _ in media_offset]
        #     flex_mask_block = []

        # 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,
                    image_embeds,
                    media_offset,
                    past_key_values,
                    output_attentions,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    image_embeds=image_embeds,
                    media_offset=media_offset,
                    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 HyperQwen2ForCausalLM(Qwen2PreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = HyperQwen2Model(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

    @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,
        image_embeds=None,
        media_offset=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

        # # media_offset to 3.5 format
        # if media_offset is not None:
        #     bs = media_offset.shape[0]
        #     pad_media_offset = torch.cat([torch.zeros(media_offset.shape[0], 1,device=media_offset.device, dtype=media_offset.dtype), media_offset], dim=1)
        #     pad_media_offset = (pad_media_offset[:,1:] - pad_media_offset[:,:-1]).nonzero()
        #     media_offset = [[] for bi in range(bs)]
        #     for i, (bi, li) in enumerate(pad_media_offset):
        #         media_offset[bi].append(li)


        # 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,
            image_embeds=image_embeds,
            media_offset=media_offset,
            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)
        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)

        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
    ):
        # 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[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                'image_embeds': kwargs.get('image_embeds'),
                'media_offset': kwargs.get('media_offset'),
            }
        )
        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