projects/ST/models/models_modeling_qwen2mm_mmrope.py

import inspect
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from typing import Optional, List, Union, Tuple
from transformers import Qwen2Model, Qwen2ForCausalLM
from transformers.utils import logging, is_torchdynamo_compiling
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.cache_utils import Cache, DynamicCache, StaticCache, SlidingWindowCache
from transformers.configuration_utils import PretrainedConfig

from transformers.modeling_outputs import (
    CausalLMOutputWithPast,
    BaseModelOutputWithPast,
)
from transformers.modeling_attn_mask_utils import (
    _prepare_4d_causal_attention_mask,
    _prepare_4d_causal_attention_mask_for_sdpa,
)
from transformers.models.qwen2.modeling_qwen2 import (
    Qwen2RMSNorm,
    Qwen2MLP
)

from transformers import AutoTokenizer, GenerationConfig
from transformers.utils import (
    add_code_sample_docstrings,
    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,
)

logger = logging.get_logger(__name__)

if is_flash_attn_2_available():
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa

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


# Impl. for transformers==4.42.0
class Qwen2MMConfig(PretrainedConfig):
    model_type = "qwen"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
            self,
            vocab_size=151936,
            hidden_size=4096,
            intermediate_size=22016,
            num_hidden_layers=32,
            num_attention_heads=32,
            num_key_value_heads=32,
            hidden_act="silu",
            max_position_embeddings=32768,
            initializer_range=0.02,
            rms_norm_eps=1e-6,
            use_cache=True,
            tie_word_embeddings=False,
            rope_theta=10000.0,
            use_sliding_window=False,
            sliding_window=4096,
            max_window_layers=28,
            attention_dropout=0.0,
            vision_patch_size=32,
            **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.use_sliding_window = use_sliding_window
        self.sliding_window = sliding_window
        self.max_window_layers = max_window_layers

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

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.vision_patch_size = vision_patch_size

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


class Qwen2mmMultimodalRotaryEmbedding(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)
        self.original_inv_freq = self.inv_freq
        self.msection = [0, dim // 4, dim // 4]

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

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


def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section=[0, 32, 32], 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


# 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)


# 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,
    )

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: Qwen2MMConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.attention_dropout = config.attention_dropout

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.rotary_emb = Qwen2mmMultimodalRotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
    ) -> 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)

        if position_embeddings is None:
            logger.warning_once(
                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
                "removed and `position_embeddings` will be mandatory."
            )
            cos, sin = self.rotary_emb(value_states, position_ids)
        else:
            cos, sin = position_embeddings
        query_states, key_states = apply_multimodal_rotary_pos_emb(query_states, key_states, cos, sin, mrope_section=self.rotary_emb.msection)

        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

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

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

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

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


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

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

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ):
        bsz, q_len, _ = hidden_states.size()

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

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

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

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

        # change to following:
        cos, sin = self.rotary_emb(value_states, position_ids)

        query_states, key_states = apply_multimodal_rotary_pos_emb(query_states, key_states, cos, sin, mrope_section=self.rotary_emb.msection)

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

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

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

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

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

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

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

            cache_kwargs = {"sin": sin, "cos": cos, "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)

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

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

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

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

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

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

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

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

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

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

        return attn_output

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

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

        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)

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

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

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


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

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

        bsz, q_len, _ = hidden_states.size()

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

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

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

        query_states, key_states = apply_multimodal_rotary_pos_emb(query_states, key_states, cos, sin, mrope_section=self.rotary_emb.msection)

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


class Qwen2DecoderLayer(nn.Module):
    def __init__(self, config: Qwen2MMConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

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

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        **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.
            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,
        )
        hidden_states = residual + hidden_states

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

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs
    

class MultimodalQwen2Model(Qwen2Model):

    def __init__(self, config: Qwen2MMConfig):
        super(Qwen2Model, self).__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

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

        self.gradient_checkpointing = False

        # === Vision Patches ===
        self.vis_embed = nn.Linear(
            config.vision_patch_size * config.vision_patch_size * 3,
            config.hidden_size,
            bias=False,
        )
        # === Vision Patches ===

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

    def forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            vision_patch_indices: torch.LongTensor = None,  # (batch_size, seq_length), "-1" for text token
            vision_patches: torch.FloatTensor = None,  # (n_patches, 32 * 32 * 3)
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            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 cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
            )

        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

        use_legacy_cache = False
        if use_cache and not isinstance(past_key_values, Cache):
            use_legacy_cache = True
            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            logger.warning_once(
                "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
                "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
            )

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

            if vision_patch_indices is not None:
                assert (
                        vision_patch_indices.shape == input_ids.shape
                ), "vision_patch_indices and input_ids should have the same shape"

            # === Handle vision patches ===
            if vision_patches is not None and (vision_patch_indices > -1).any().item():
                assert vision_patch_indices is not None, "HF QwenMM model requires vision_patch_indices for vision_patches input."
                vision_embeds = self.vis_embed(vision_patches)  # (n_patches, hidden_size)
                vision_embeds = torch.cat(
                    [
                        vision_embeds,
                        torch.zeros(1, self.config.hidden_size).to(
                            vision_embeds.device
                        ),  # add a dummy token (for text)
                    ],
                )  # (n_patches + 1, hidden_size)
                # arrange embeddings according to vision_patch_indices
                # - text tokens are -1 (map to the dummy zero tensor)
                # - vision tokens are 0~n_patches (map to the corresponding vision_embeds)
                vision_embeds = vision_embeds[vision_patch_indices]  # (batch_size, seq_length, hidden_size)

                # merge vision_embeds with inputs_embeds
                inputs_embeds += vision_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 position_ids is None:
            if self.training:
                raise ValueError("Please specify position ids during training.")
            else:
                logger.warning("No position_ids detected in inference, using token cache position. Please check the correctness.")
                position_ids = cache_position.view(1,1,-1).expand(3, inputs_embeds.shape[0], -1)
        elif position_ids.ndim == 2:
            position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)

        # hack: save max_postion_index for inference
        if past_key_values is not None:
            if not isinstance(past_key_values, Cache):
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
                return_legacy_cache = False
            past_key_values.max_position_index = position_ids.max().item()


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

        hidden_states = inputs_embeds

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

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

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

            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,
        )

    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114

        if self.config._attn_implementation == "flash_attention_2":
            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)

        # 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 and not output_attentions:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                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]
        if using_static_cache:
            target_length = past_key_values.get_max_length()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        if attention_mask is not None and attention_mask.dim() == 4:
            # change to invert form for the following:
            if attention_mask.dtype == torch.bool:
                zero_weight_mask = (attention_mask == True)
                _attention_mask = torch.full(attention_mask.size(), min_dtype, dtype=dtype, device=device)
                _attention_mask.masked_fill_(zero_weight_mask, 0.)
                attention_mask = _attention_mask

            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
            if attention_mask.max() != 0:
                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
            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(input_tensor.shape[0], 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
                )
        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            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


class Qwen2MMmropeForCausalLM(Qwen2ForCausalLM):

    def __init__(self, config: Qwen2MMConfig):
        super().__init__(config)
        self.model = MultimodalQwen2Model(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 forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            vision_patch_indices: torch.LongTensor = None,  # (batch_size, seq_length), "-1" for text token
            vision_patches: torch.FloatTensor = None,  # (n_patches, 32 * 32 * 3)
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            cache_position: Optional[torch.LongTensor] = None,
            num_logits_to_keep: int = 0,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            vision_patch_indices=vision_patch_indices,
            vision_patches=vision_patches,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        hidden_states = outputs[0]
        if labels is None and not is_torchdynamo_compiling():
            logger.warning_once(
                "Starting from v4.46, the `logits` model output will have the same type as the model (except at train time, where it will always be FP32)"
            )
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        # TODO: remove the float() operation in v4.46
        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :]).float()

        loss = None
        if labels is not None:
            # Upcast to float if we need to compute the loss to avoid potential precision issues
            logits = logits.float()
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = nn.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,
        )

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.prepare_inputs_for_generation
    def prepare_inputs_for_generation(
            self,
            input_ids,
            past_key_values=None,
            attention_mask=None,
            inputs_embeds=None,
            cache_position=None,
            use_cache=True,
            **kwargs,
    ):
        assert input_ids.size(0) == 1, "Currently only support bsz=1"

        vision_patches = kwargs.get("vision_patches", None)
        vision_patch_indices = kwargs.get("vision_patch_indices", None)

        has_vision_inp = False
        if vision_patches is not None and vision_patch_indices is not None:
            has_vision_inp = True
            # make vision_patch_indices to be the same shape as input_ids by padding -1
            _padding = torch.full_like(input_ids, -1, dtype=vision_patch_indices.dtype)
            _padding[:, : vision_patch_indices.shape[1]] = vision_patch_indices
            vision_patch_indices = _padding

        max_position_index = -1
        past_length = 0

        ori_input_ids = input_ids   # input_ids received by original user input, or (original user input + generated token) afterwards

        # Omit tokens covered by past_key_values
        if past_key_values is not None:
            # Past key values are always initialized with a `Cache` object -> no need for if-else anymore
            past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length()
            max_cache_length = (
                torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
                if past_key_values.get_max_length() is not None
                else None
            )
            cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)

            # 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):]
                if has_vision_inp:
                    vision_patch_indices = vision_patch_indices[:, -(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:]
                if has_vision_inp:
                    vision_patch_indices = vision_patch_indices[:, 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 past_key_values is not None and position_ids is None:
            max_position_index = getattr(past_key_values, "max_position_index", max_position_index)
            position_ids = torch.arange(max_position_index+1, max_position_index+1+input_ids.size(1)).to(input_ids)
            position_ids = position_ids.unsqueeze(0).expand(3, input_ids.size(0), -1)
        elif 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)
            position_ids = position_ids.unsqueeze(0).expand(3, input_ids.size(0), -1)
            assert position_ids.size(-1) == input_ids.size(-1)
        elif position_ids is not None:
            assert position_ids.size(0) == 3
            if position_ids.size(-1) != ori_input_ids.size(-1):
                # 1 :for the case with tokens added during generation
                assert position_ids.size(-1) <= ori_input_ids.size(-1)
                max_position_index = position_ids.max().item()
                new_pos = torch.arange(
                    max_position_index+1, 
                    max_position_index+1+ori_input_ids.size(-1)-position_ids.size(-1)
                ).unsqueeze(0).expand(3, position_ids.size(1), -1).to(position_ids)
                position_ids = torch.cat([position_ids, new_pos], dim=-1)
                assert position_ids.size(-1) == ori_input_ids.size(-1)
                position_ids = position_ids[:,:,-input_ids.size(-1):]
            elif position_ids.size(-1) == ori_input_ids.size(-1) and position_ids.size(-1) != input_ids.size(-1):
                # 2: for the case with some pre-computed cache (customized attention mask), and the rest with causal attention
                position_ids = position_ids[:,:,-input_ids.size(-1):]

        assert position_ids is not None, "For the mmrope model, please make sure `position_ids` is passed to the generate function."
        # print("input ids shape:{} , position_ids: {}.".format(input_ids.shape, position_ids.squeeze(-1).squeeze(-1)))

        # crop the attention_mask to sliding window size during decode phase if using SlidingWindowCache
        if (
            past_length > 0
            and attention_mask is not None
            and isinstance(past_key_values, SlidingWindowCache)
            and attention_mask.shape[1] > past_key_values.max_cache_len
        ):
            attention_mask = attention_mask[:, -past_key_values.max_cache_len :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_length == 0:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1]
        if cache_position is None:
            cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device)
        elif use_cache:
            cache_position = cache_position[-input_length:]
        
        if vision_patch_indices is not None:
            assert vision_patch_indices.shape == input_ids.shape

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
                "vision_patch_indices": vision_patch_indices,
                "vision_patches": vision_patches,
            }
        )
        return model_inputs


def get_transformer_and_tokenizer(model_path, tokenizer_path):
    model = Qwen2MMmropeForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, use_cache=False, attn_implementation="sdpa")
    tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, padding="right", trust_remote_mode=True)
    tokenizer.vis_beg_tok = "<vision>"
    tokenizer.vis_patch_tok = "<vpatch>"
    tokenizer.vis_rsep_tok = "<vrow_sep>"
    tokenizer.vis_frm_tok = "<vframe_sep>"
    tokenizer.vis_end_tok = "</vision>"
    tokenizer.vis_cls_tok = "<|vis_cls|>"

    tokenizer.vis_beg_tok_id = tokenizer.convert_tokens_to_ids("<vision>")
    tokenizer.vis_patch_tok_id = tokenizer.convert_tokens_to_ids("<vpatch>")
    tokenizer.vis_rsep_tok_id = tokenizer.convert_tokens_to_ids("<vrow_sep>")
    tokenizer.vis_frm_tok_id = tokenizer.convert_tokens_to_ids("<vframe_sep>")
    tokenizer.vis_end_tok_id = tokenizer.convert_tokens_to_ids("</vision>")
    tokenizer.vis_cls_tok_id = tokenizer.convert_tokens_to_ids("<|vis_cls|>")
    
    return model, tokenizer