Ming_Uni/modeling_qwen2_native.py

import math

import torch
import torch.nn as nn
from transformers.models.qwen2.modeling_qwen2 import (
    Qwen2MLP,
    Qwen2RMSNorm,
    Qwen2PreTrainedModel,
    rotate_half,
    repeat_kv,
    QWEN2_START_DOCSTRING,
    QWEN2_INPUTS_DOCSTRING,
    Qwen2RotaryEmbedding,
    apply_rotary_pos_emb
)

from IPython import embed

from transformers.cache_utils import Cache, SlidingWindowCache, StaticCache
from .modeling_rope_utils import ROPE_INIT_FUNCTIONS, rope_config_validation
from transformers.models.qwen2.configuration_qwen2 import Qwen2Config

from dataclasses import dataclass
from typing import List, Optional, Tuple, Union, Dict, Any
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_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_outputs import BaseModelOutputWithPast, ModelOutput

if is_flash_attn_2_available():
    from transformers.modeling_flash_attention_utils import _flash_attention_forward
else:
    flash_attn_varlen_func = None

_CONFIG_FOR_DOC = "Qwen2Config"
logger = logging.get_logger(__name__)

@dataclass
class Bailing2CausalLMOutputWithPast(ModelOutput):
    """
    Base class for Bailing2 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
    last_hidden_state: Optional[torch.FloatTensor] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None

class Qwen2_5_VLRotaryEmbedding(nn.Module):
    def __init__(self, config: Qwen2Config, device=None):
        super().__init__()
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_scaling = config.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"})
            self.rope_type = self.rope_scaling["rope_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.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):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block. In contrast to other models, Qwen2 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)

def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section=[16, 24, 24], 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

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

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

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.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)

        self.use_llm_3drope = config.use_llm_3drope
        if self.use_llm_3drope:
            self.rotary_emb = Qwen2_5_VLRotaryEmbedding(config=config)
        else:
            self.rotary_emb = Qwen2RotaryEmbedding(config=config)

    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,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ) -> 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, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        if self.use_llm_3drope:
            query_states, key_states = apply_multimodal_rotary_pos_emb(
                query_states, key_states, cos, sin,
                mrope_section=self.rope_scaling["mrope_section"],
            )
        else:
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

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

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

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

        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        # Fix precision issues in Qwen2-VL float16 inference
        # Replace inf values with zeros in attention weights to prevent NaN propagation
        if query_states.dtype == torch.float16:
            attn_weights = torch.where(torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights)

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

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

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

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

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

    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,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ):
        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, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        # Because the input can be padded, the absolute sequence length depends on the max position id.
        cos, sin = position_embeddings
        if self.use_llm_3drope:
            query_states, key_states = apply_multimodal_rotary_pos_emb(
                query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
            )
        else:
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # 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 (
                self.config.use_sliding_window
                and getattr(self.config, "sliding_window", None) is not None
                and self.layer_idx >= self.config.max_window_layers
        ):
            sliding_window = self.config.sliding_window
        else:
            sliding_window = None

        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            sliding_window=sliding_window,
            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

class Qwen2SdpaAttention(Qwen2Attention):
    """
    Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from Qwen2Attention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            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,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
            )

        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, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        if self.use_llm_3drope:
            query_states, key_states = apply_multimodal_rotary_pos_emb(
                query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
            )
        else:
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # 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)

        causal_mask = attention_mask
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]

        # 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()

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        # 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 = True if causal_mask is None and q_len > 1 else False

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=causal_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            is_causal=is_causal,
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value

QWEN2_5_ATTENTION_CLASSES = {
    "eager": Qwen2Attention,
    "flash_attention_2": Qwen2FlashAttention2,
    "sdpa": Qwen2SdpaAttention,
}

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

        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = QWEN2_5_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,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **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,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
        )
        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

@add_start_docstrings(
    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
    QWEN2_START_DOCSTRING,
)
class Qwen2Model(Qwen2PreTrainedModel):
    def __init__(self, config: Qwen2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

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

        self.use_llm_3drope = config.use_llm_3drope
        if self.use_llm_3drope:
            self.rotary_emb = Qwen2_5_VLRotaryEmbedding(config=config)
        else:
            self.rotary_emb = Qwen2RotaryEmbedding(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 prompt_wrap(self, input_ids, query_embeds_visual=None, query_embeds_audio=None, target_embeds=None):
        inputs_embeds = self.embed_tokens(input_ids)
        if query_embeds_visual is None and query_embeds_audio is None and target_embeds is None:
            return inputs_embeds

        if query_embeds_visual is not None:
            inputs_embeds = inputs_embeds.to(dtype=query_embeds_visual.dtype, device=query_embeds_visual.device)
            image_mask = input_ids == self.config.image_patch_token
            query_embeds_visual = query_embeds_visual.view(-1, query_embeds_visual.shape[-1])
            try:
                inputs_embeds[image_mask] = query_embeds_visual
            except Exception as e:
                temp_embeds = torch.zeros_like(inputs_embeds[image_mask]).to(dtype=inputs_embeds.dtype,
                    device=inputs_embeds.device)
                inputs_embeds[image_mask] = temp_embeds
                return inputs_embeds

        if query_embeds_audio is not None:
            inputs_embeds = inputs_embeds.to(dtype=query_embeds_audio.dtype, device=query_embeds_audio.device)
            audio_mask = input_ids == self.config.audio_patch_token
            query_embeds_audio = query_embeds_audio.view(-1, query_embeds_audio.shape[-1])
            inputs_embeds[audio_mask] = query_embeds_audio

        if target_embeds is not None:
            inputs_embeds = inputs_embeds.to(dtype=target_embeds.dtype, device=target_embeds.device)
            target_mask = input_ids == self.config.gen_image_patch_token
            target_embeds = target_embeds.view(-1, target_embeds.shape[-1])
            inputs_embeds[target_mask] = target_embeds

        return inputs_embeds

    def prompt_wrap_vision(self, input_ids, inputs_embeds, vision_embeds, image_token_id=None):
        if vision_embeds is None or input_ids is None:
            return inputs_embeds
        if len(vision_embeds.shape) == 3:
            vision_embeds = vision_embeds.reshape(-1, vision_embeds.shape[-1])
        self.config.image_token_id = image_token_id if image_token_id is not None else self.config.image_patch_token
        n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
        n_image_features = vision_embeds.shape[0]
        if n_image_tokens != n_image_features:
            raise ValueError(
                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
            )
        image_mask = (
            (input_ids == self.config.image_token_id)
            .unsqueeze(-1)
            .expand_as(inputs_embeds)
            .to(inputs_embeds.device)
        )
        #if torch.distributed.get_rank() == 0:
        #    embed()
        #torch.distributed.barrier()


        image_embeds = vision_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
        inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
        return inputs_embeds

    def prompt_wrap_audio(self, input_ids, inputs_embeds, audio_embeds, audio_token_id=None):
        if audio_embeds is None or input_ids is None:
            return inputs_embeds
        if len(audio_embeds.shape) == 3:
            audio_embeds = audio_embeds.reshape(-1, audio_embeds.shape[-1])
        self.config.audio_token_id = audio_token_id if audio_token_id is not None else self.config.audio_patch_token
        n_audio_tokens = (input_ids == self.config.audio_token_id).sum().item()
        n_audio_features = audio_embeds.shape[0]
        if n_audio_tokens != n_audio_features:
            raise ValueError(
                f"Audio features and audio tokens do not match: tokens: {n_audio_tokens}, features {n_audio_features}"
            )
        audio_mask = (
            (input_ids == self.config.audio_token_id)
            .unsqueeze(-1)
            .expand_as(inputs_embeds)
            .to(inputs_embeds.device)
        )
        audio_embeds = audio_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
        inputs_embeds = inputs_embeds.masked_scatter(audio_mask, audio_embeds)
        return inputs_embeds

    def prompt_wrap_navit(self, input_ids, query_embeds_image=None, query_embeds_video=None, query_embeds_audio=None,
        target_embeds=None):
        inputs_embeds = self.embed_tokens(input_ids)
        if query_embeds_image is None and query_embeds_video is None and query_embeds_audio is None and target_embeds is None:
            return inputs_embeds
        if query_embeds_image is not None:
            inputs_embeds = self.prompt_wrap_vision(input_ids, inputs_embeds, query_embeds_image)
        if query_embeds_video is not None:
            inputs_embeds = self.prompt_wrap_vision(input_ids, inputs_embeds, query_embeds_video)
        if query_embeds_audio is not None:
            inputs_embeds = self.prompt_wrap_audio(input_ids, inputs_embeds, query_embeds_audio)
        return inputs_embeds

    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,
        query_embeds_image: Optional[torch.Tensor] = None,
        query_embeds_video: Optional[torch.Tensor] = None,
        query_embeds_audio: Optional[torch.Tensor] = None,
        target_embeds: Optional[torch.Tensor] = None,
        img_gen_embeds: Optional[torch.Tensor] = 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,
        image_grid_thw: Optional[torch.Tensor] = None,
        image_grid_thw_video: Optional[torch.Tensor] = None,
        cache_position: Optional[torch.LongTensor] = 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

        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

        if inputs_embeds is None:
            if (
                    query_embeds_image is None
                    and query_embeds_video is None
                    and query_embeds_audio is None
                    and target_embeds is None
            ) or input_ids.size(1) == 1:  # only text_ids
                inputs_embeds = self.embed_tokens(input_ids.clip(0, self.embed_tokens.weight.shape[0] - 1))
            else:
                if image_grid_thw is None and image_grid_thw_video is None:
                    inputs_embeds = self.prompt_wrap(
                        input_ids.clip(0, self.embed_tokens.weight.shape[0] - 1), query_embeds_image,
                        query_embeds_audio, target_embeds  # noqa
                    )
                else:
                    # print("query_embeds_image: ", query_embeds_image.shape)
                    # print("image_grid_thw:", image_grid_thw, image_grid_thw.shape)
                    inputs_embeds = self.prompt_wrap_navit(
                        input_ids.clip(0, self.embed_tokens.weight.shape[0] - 1), query_embeds_image,
                        query_embeds_video, query_embeds_audio, target_embeds)

        if img_gen_embeds is not None:
            gen_length = img_gen_embeds.shape[1]
            inputs_embeds[:, -gen_length:] = img_gen_embeds

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        if self.use_llm_3drope:
            # the hard coded `3` is for temporal, height and width.
            if position_ids is None:
                position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
            elif position_ids.dim() == 2:
                position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)
        else:
            if position_ids is None:
                position_ids = cache_position.unsqueeze(0)

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )

        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,
                    causal_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    position_embeddings,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position,
                    position_embeddings=position_embeddings,
                )

            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,
        )

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and past_key_values is not None:
                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
                if is_padding_right:
                    logger.warning_once(
                        "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 attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)
        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if (
                self.config._attn_implementation == "sdpa"
                and not (using_static_cache or using_sliding_window_cache)
                and not output_attentions
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                    attention_mask,
                    inputs_embeds=input_tensor,
                    past_key_values_length=past_seen_tokens,
                    sliding_window=self.config.sliding_window,
                    is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        # SlidingWindowCache or StaticCache
        if using_sliding_window_cache or using_static_cache:
            # target_length = past_key_values.get_max_cache_shape()
            target_length = past_key_values.get_max_length()
        # DynamicCache or no cache
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            min_dtype=min_dtype,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
                self.config._attn_implementation == "sdpa"
                and attention_mask is not None
                and attention_mask.device.type in ["cuda", "xpu"]
                and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        min_dtype: float,
        cache_position: torch.Tensor,
        batch_size: int,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to plcae the 4D attention mask on.
            min_dtype (`float`):
                The minimum value representable with the dtype `dtype`.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask

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

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

        self.use_llm_3drope = config.use_llm_3drope
        if self.use_llm_3drope:
            self.config.rope_scaling = {"type": "mrope", "mrope_section": [16, 24, 24]}

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

        self.audio_vocab_size = config.audio_vocab_size
        self.audio_id_shift = config.audio_id_shift

        # 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 audio_decoder_sample(self, logits, topk=10, filter_value=-float("Inf")):
        """
        - logits: size(batch, audio_vocab_size)

        Return
        - token_id: int
        """
        assert logits.dim() == 2 and logits.size(1) == self.config.audio_vocab_size
        indices_to_remove = logits < torch.topk(logits, topk)[0][..., -1, None]
        logits[indices_to_remove] = filter_value
        token_id = torch.multinomial(torch.softmax(logits, dim=-1), num_samples=1)
        return token_id

    def get_rope_index(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = 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 modern 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:
                Temporal (Time): 3 patches, representing different segments of the video in time.
                Height: 2 patches, dividing each frame vertically.
                Width: 2 patches, dividing each frame horizontally.
                We also have some important parameters:
                fps (Frames Per Second): The video's frame rate, set to 1. This means one frame is processed each second.
                tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens" are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens per second. So each second of the video will be represented with 25 separate time points. It essentially defines the temporal granularity.
                temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames.
                interval: The step size for the temporal position IDs, calculated as tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50. This means that each temporal patch will be have a difference of 50 in the temporal position IDs.
                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, 50, 50, 50, 50, 100, 100, 100, 100]
                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: [101, 102, 103, 104, 105]
                text height position_ids: [101, 102, 103, 104, 105]
                text width position_ids: [101, 102, 103, 104, 105]
                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.
            second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
                The time interval (in seconds) for each grid along the temporal dimension in the 3D position 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**.

        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
        image_token_id = self.config.image_patch_token
        video_token_id = self.config.video_patch_token
        image_start_token_id = self.config.image_start_token
        video_start_token_id = self.config.video_start_token
        use_abs_time_pos = second_per_grid_ts is not None

        mrope_position_deltas = []
        if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
            total_input_ids = input_ids
            if attention_mask is None:
                attention_mask = torch.ones_like(total_input_ids)
            position_ids = torch.ones(
                3,
                input_ids.shape[0],
                input_ids.shape[1],
                dtype=input_ids.dtype,
                device=input_ids.device,
            )
            image_index, video_index = 0, 0
            attention_mask = attention_mask.to(total_input_ids.device)
            for i, input_ids in enumerate(total_input_ids):
                input_ids = input_ids[attention_mask[i] == 1]
                image_nums, video_nums = 0, 0
                if image_grid_thw is not None:
                    vision_start_indices = torch.argwhere(input_ids == image_start_token_id).squeeze(1)
                    vision_tokens = input_ids[vision_start_indices + 1]
                    image_nums = (vision_tokens == image_token_id).sum()
                if video_grid_thw is not None:
                    vision_start_indices = torch.argwhere(input_ids == video_start_token_id).squeeze(1)
                    vision_tokens = input_ids[vision_start_indices + 1]
                    video_nums = (vision_tokens == video_token_id).sum()

                input_tokens = input_ids.tolist()
                llm_pos_ids_list: list = []
                st = 0
                remain_images, remain_videos = image_nums, video_nums
                for _ in range(image_nums + video_nums):
                    if image_token_id in input_tokens and remain_images > 0:
                        ed_image = input_tokens.index(image_token_id, st)
                    else:
                        ed_image = len(input_tokens) + 1
                    if video_token_id in input_tokens and remain_videos > 0:
                        ed_video = input_tokens.index(video_token_id, st)
                    else:
                        ed_video = len(input_tokens) + 1
                    if ed_image < ed_video:
                        t, h, w = (
                            image_grid_thw[image_index][0],
                            image_grid_thw[image_index][1],
                            image_grid_thw[image_index][2],
                        )
                        second_per_grid_t = 0
                        image_index += 1
                        remain_images -= 1
                        ed = ed_image
                    else:
                        t, h, w = (
                            video_grid_thw[video_index][0],
                            video_grid_thw[video_index][1],
                            video_grid_thw[video_index][2],
                        )
                        if second_per_grid_ts is not None:
                            second_per_grid_t = second_per_grid_ts[video_index]
                        else:
                            second_per_grid_t = 1.0
                        video_index += 1
                        remain_videos -= 1
                        ed = ed_video
                    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)

                    range_tensor = torch.arange(llm_grid_t).view(-1, 1)
                    expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)
                    if use_abs_time_pos:
                        time_tensor = expanded_range * second_per_grid_t * self.config.tokens_per_second
                        time_tensor_long = time_tensor.long()
                    else:
                        time_tensor_long = expanded_range.long()
                    t_index = time_tensor_long.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)
                mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
            mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
            return position_ids, mrope_position_deltas
        else:
            if attention_mask is not None:
                position_ids = attention_mask.long().cumsum(-1) - 1
                position_ids.masked_fill_(attention_mask == 0, 1)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
                max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
                mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
            else:
                position_ids = (
                    torch.arange(input_ids.shape[1], device=input_ids.device)
                    .view(1, 1, -1)
                    .expand(3, input_ids.shape[0], -1)
                )
                mrope_position_deltas = torch.zeros(
                    [input_ids.shape[0], 1],
                    device=input_ids.device,
                    dtype=input_ids.dtype,
                )
            return position_ids, mrope_position_deltas

    def _update_model_kwargs_for_generation(
        self,
        outputs: ModelOutput,
        model_kwargs: Dict[str, Any],
        is_encoder_decoder: bool = False,
        num_new_tokens: int = 1,
    ) -> Dict[str, Any]:
        model_kwargs = super()._update_model_kwargs_for_generation(
            outputs=outputs,
            model_kwargs=model_kwargs,
            is_encoder_decoder=is_encoder_decoder,
            num_new_tokens=num_new_tokens,
        )

        if getattr(outputs, "rope_deltas", None) is not None:
            model_kwargs["rope_deltas"] = outputs.rope_deltas

        return model_kwargs

    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Bailing2CausalLMOutputWithPast, 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,
        query_embeds_image: Optional[torch.Tensor] = None,
        query_embeds_video: Optional[torch.Tensor] = None,
        query_embeds_audio: Optional[torch.Tensor] = None,
        target_embeds: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        img_gen_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        reduction: Optional[str] = "mean",
        weights=None,
        is_pretrain=False,
        image_grid_thw: Optional[torch.LongTensor] = None,
        image_grid_thw_video: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        is_audio_generation_mode=False,
        no_image_end_prediction=False,
    ) -> Union[Tuple, Bailing2CausalLMOutputWithPast]:
        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]`.

            logits_to_keep (`int` or `torch.Tensor`, *optional*):
                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
                This is useful when using packed tensor format (single dimension for batch and sequence length).

        Returns:

        Example:

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

        >>> model = Qwen2ForCausalLM.from_pretrained("meta-qwen2/Qwen2-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-qwen2/Qwen2-2-7b-hf")

        >>> 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

        ignore_flag = False
        if self.use_llm_3drope:
            # update position_ids for llm_3drope
            if position_ids is None and input_ids is not None:
                # try:
                #     position_ids, _ = self.get_rope_index(input_ids, image_grid_thw, image_grid_thw_video,
                #         attention_mask)
                # except Exception as e:
                #     position_ids, _ = self.get_rope_index(input_ids, attention_mask=attention_mask)
                #     ignore_flag = True
                position_ids, _ = self.get_rope_index(input_ids, image_grid_thw, image_grid_thw_video, attention_mask)

        #embed()
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            query_embeds_image=query_embeds_image,
            query_embeds_video=query_embeds_video,
            query_embeds_audio=query_embeds_audio,
            target_embeds=target_embeds,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            img_gen_embeds=img_gen_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            image_grid_thw=image_grid_thw,
            image_grid_thw_video=image_grid_thw_video,
            cache_position=cache_position,
        )

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

        if is_audio_generation_mode is True:
            need_replace = torch.argmax(logits[:, -1, :], -1) >= self.audio_id_shift
            next_audio_token_logits_for_generation = logits[:, -1, self.audio_id_shift:]
            next_audio_token_for_generation = (
                    self.audio_decoder_sample(next_audio_token_logits_for_generation) + self.audio_id_shift).view(
                -1)
            logits[torch.tensor(range(logits.size(0)), device=logits.device)[need_replace], -1,
            next_audio_token_for_generation[need_replace]] = 99999

        loss = None
        assert labels is None  
        

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

        return Bailing2CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=rope_deltas,
            last_hidden_state=outputs.last_hidden_state,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        query_embeds_image=None,
        query_embeds_video=None,
        query_embeds_audio=None,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        image_grid_thw=None,
        image_grid_thw_video=None,
        second_per_grid_ts=None,
        is_audio_generation_mode=False,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        # Exception 3: If input_embeds are passed then slice it through `cache_position`, to keep only the unprocessed tokens and
        # generate the first token for each sequence. Later use the generated Input ids for continuation.
        if past_key_values is not None:
            if inputs_embeds is not None:
                input_ids = input_ids[:, -cache_position.shape[0]:]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        img_gen_embeds = None

        rope_deltas = kwargs.get("rope_deltas", None)
        if attention_mask is not None and position_ids is None:
            if self.use_llm_3drope:
                if cache_position is None or (cache_position is not None and cache_position[0] == 0):
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids, image_grid_thw, image_grid_thw_video, attention_mask
                    )
                else:
                    batch_size, seq_length = input_ids.shape
                    delta = (
                        cache_position[0] + rope_deltas if cache_position is not None and rope_deltas is not None else 0
                    )
                    position_ids = torch.arange(seq_length, device=input_ids.device)
                    position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                    position_ids = position_ids.add(delta)
                    position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
            else:
                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]:]

                    # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
                    position_ids = position_ids.clone(memory_format=torch.contiguous_format)

            if cache_position[0] != 0:
                query_embeds_image = None
                query_embeds_video = None
                query_embeds_audio = None

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and len(cache_position) == inputs_embeds.shape[1]:
            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
        else:
            model_inputs = {"input_ids": input_ids, "inputs_embeds": None}

        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
            if model_inputs["inputs_embeds"] is not None:
                batch_size, sequence_length, _ = inputs_embeds.shape
                device = inputs_embeds.device
            else:
                batch_size, sequence_length = input_ids.shape
                device = input_ids.device

            attention_mask = self.model._prepare_4d_causal_attention_mask_with_cache_position(
                attention_mask,
                sequence_length=sequence_length,
                target_length=past_key_values.get_max_cache_shape(),
                dtype=self.lm_head.weight.dtype,
                device=device,
                cache_position=cache_position,
                batch_size=batch_size,
            )

        model_inputs.update(
            {
                "position_ids": position_ids,
                "query_embeds_image": query_embeds_image,
                "query_embeds_video": query_embeds_video,
                "query_embeds_audio": query_embeds_audio,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "img_gen_embeds": img_gen_embeds,
                "image_grid_thw": image_grid_thw,
                "image_grid_thw_video": image_grid_thw_video,
                "cache_position": cache_position,
                "rope_deltas": rope_deltas,
                "second_per_grid_ts": second_per_grid_ts,
                "is_audio_generation_mode": is_audio_generation_mode,
            }
        )
        return model_inputs