modeling_youtu_vl.py

# coding=utf-8
# Copyright 2026 Tencent Youtu lab, DeepSeek-AI and The HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import math
import os
from functools import partial
from typing import Callable, Optional, Tuple, Union, List, Any, Dict

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

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, StaticCache
from transformers.generation import GenerationMixin
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    can_return_tuple,
    is_torch_flex_attn_available,
    logging,
    replace_return_docstrings,
    is_flash_attn_2_available,
)
from transformers.utils.deprecation import deprecate_kwarg
from .configuration_youtu_vl import YoutuVLConfig

from .modeling_siglip2 import Siglip2VisionModel, Siglip2VisionEmbeddings
from .configuration_siglip2 import Siglip2VisionConfig

if is_torch_flex_attn_available():
    from torch.nn.attention.flex_attention import BlockMask
    from transformers.integrations.flex_attention import make_flex_block_causal_mask

is_aiter_available = False

if is_flash_attn_2_available():
    try:
        from aiter import flash_attn_varlen_func
        is_aiter_available = True
    except ImportError:
        from flash_attn import flash_attn_varlen_func
else:
    flash_attn_varlen_func = None
    
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "YoutuVLConfig"

class YoutuRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype

        hidden_states = hidden_states.to(torch.float32)

        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


class YoutuRotaryEmbedding(nn.Module):
    def __init__(self, config: YoutuVLConfig, device=None):
        super().__init__()
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

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

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

    @torch.no_grad()
    @dynamic_rope_update  
    def forward(self, x, position_ids):
        """
        Compute rotary positional embeddings.
        
        Args:
            x (torch.Tensor): Input tensor, shape (batch_size, seq_len, feature_dim)
            position_ids (torch.LongTensor): Position indices, shape (batch_size, seq_len)

        Returns:
            Tuple of (cos, sin) tensors for rotary embedding
        """
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

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


class YoutuMLP(nn.Module):
    def __init__(self, config, hidden_size=None, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size

        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


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_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

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


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


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


def apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    r"""
    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)

    b, h, s, d = q.shape
    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)

    b, h, s, d = k.shape
    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)

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


def yarn_get_mscale(scale=1, mscale=1):
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


class YoutuMLAttention(nn.Module):
    """
    Multi-latent attention from 
    'DeepSeek-V2: A Strong, Economical, 
    and Efficient Mixture-of-Experts Language Model'paper
    """
    
    def __init__(self, config: YoutuVLConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.num_key_value_groups = 1  # needed for eager attentions
        self.attention_dropout = config.attention_dropout
        self.num_heads = config.num_attention_heads
        self.rope_theta = config.rope_theta
        self.q_lora_rank = config.q_lora_rank
        self.qk_rope_head_dim = config.qk_rope_head_dim
        self.kv_lora_rank = config.kv_lora_rank
        self.v_head_dim = config.v_head_dim
        self.qk_nope_head_dim = config.qk_nope_head_dim
        self.qk_head_dim = config.qk_head_dim
        self.flash_att_sliding_window = config.flash_att_sliding_window
        self.is_causal = True

        if self.q_lora_rank is None:
            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
        else:
            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
            self.q_a_layernorm = YoutuRMSNorm(config.q_lora_rank)
            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)

        self.kv_a_proj_with_mqa = nn.Linear(
            config.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=config.attention_bias,
        )
        self.kv_a_layernorm = YoutuRMSNorm(self.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
        )

        self.o_proj = nn.Linear(
            self.num_heads * self.v_head_dim,
            config.hidden_size,
            bias=config.attention_bias,
        )

        self.scaling = self.qk_head_dim ** (-0.5)
        if self.config.rope_scaling is not None:
            mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
            scaling_factor = self.config.rope_scaling["factor"]
            if mscale_all_dim:
                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
                self.scaling = self.scaling * mscale * mscale

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        instance_length: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        batch_size, seq_length = hidden_states.shape[:-1]
        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)

        if self.q_lora_rank is None:
            q_states = self.q_proj(hidden_states).view(query_shape).transpose(1, 2)
        else:
            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states))).view(query_shape).transpose(1, 2)
        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)

        cos, sin = position_embeddings
        if self.config.rope_interleave:  # support using interleaved weights for efficiency
            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
        else:
            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)

        query_states = torch.cat((q_pass, q_rot), dim=-1)
        key_states = torch.cat((k_pass, k_rot), dim=-1)

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

        if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim:
            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
                logger.warning_once(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support"
                    "`output_attentions=True`. Falling back to 'eager attention. This warning"
                    'can be removed using the argument `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        if instance_length is None or flash_attn_varlen_func is None:
            attn_output, attn_weights = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                **kwargs,
            )
            if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim:
                attn_output = attn_output[:, :, :, : self.v_head_dim]
        else:
            instance_length = instance_length.view(-1)
            query_states = query_states.squeeze(0).transpose(0,1)
            key_states = key_states.squeeze(0).transpose(0,1)
            value_states = value_states.squeeze(0).transpose(0,1)
            max_seqlen_in_batch = instance_length.max().item()
            cu_seqlens = F.pad(torch.cumsum(instance_length, dim=0, dtype=torch.int32), (1, 0))
            if is_aiter_available:
                attn_output = flash_attn_varlen_func(query_states, key_states, value_states, cu_seqlens,
                    cu_seqlens, max_seqlen_in_batch, max_seqlen_in_batch,
                    dropout_p=0.0 if not self.training else self.attention_dropout, 
                    softmax_scale=self.scaling, 
                    causal=self.is_causal, return_lse=True)[0]
            else:
                attn_output = flash_attn_varlen_func(query_states, key_states, value_states, cu_seqlens,
                    cu_seqlens, max_seqlen_in_batch, max_seqlen_in_batch,
                    dropout_p=0.0 if not self.training else self.attention_dropout, 
                    softmax_scale=self.scaling, 
                    causal=self.is_causal)

            attn_output = attn_output.unsqueeze(0)
            attn_output = attn_output[:, :, :, : self.v_head_dim]
            attn_weights = None

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class YoutuDecoderLayer(nn.Module):
    def __init__(self, config: YoutuVLConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = YoutuMLAttention(config=config, layer_idx=layer_idx)
        self.mlp = YoutuMLP(config)
        self.input_layernorm = YoutuRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = YoutuRMSNorm(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[Cache] = None,
        output_attentions: Optional[bool] = False,
        instance_length: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        hidden_states, self_attn_weights = 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,
            instance_length=instance_length,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        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,)

        return outputs


YOUTU_VL_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`YoutuVLConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare Youtu Model outputting raw hidden-states without any specific head on top.",
    YOUTU_VL_START_DOCSTRING,
)
class YoutuPreTrainedModel(PreTrainedModel):
    config_class = YoutuVLConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["YoutuDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True
    _supports_attention_backend = True

    def init_weights(self):
        if self.config.pruned_heads:
            self.prune_heads(self.config.pruned_heads)

        if "-init" in self.name_or_path:
            self.apply(self._initialize_weights)

            for name, module in self.named_modules():
                if "o_proj" in name or "down_proj" in name:
                    scaled_std = self.config.initializer_range * (1.0 / self.config.num_hidden_layers) ** 0.5
                    module.weight.data.normal_(mean=0.0, std=scaled_std)

            self.tie_weights()

    def _init_weights(self, module):
        std = self.config.initializer_range
        embedding_std = self.config.embedding_initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=embedding_std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.Parameter):
            module.weight.data.normal_(mean=0.0, std=std)
        elif isinstance(module, YoutuRMSNorm):
            module.weight.data.fill_(1.0)


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

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

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

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

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

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

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

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

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

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

            It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

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


@add_start_docstrings(
    "The bare Youtu Model outputting raw hidden-states without any specific head on top.",
    YOUTU_VL_START_DOCSTRING,
)
class YoutuModel(YoutuPreTrainedModel):
    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.61.*"]

    def __init__(self, config: YoutuVLConfig):
        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(
            [YoutuDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = YoutuRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = YoutuRotaryEmbedding(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

    @can_return_tuple
    @add_start_docstrings_to_model_forward(YOUTU_VL_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        instance_length: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> 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

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

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
            
        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        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:
            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
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                instance_length=instance_length,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **flash_attn_kwargs,
            )
            hidden_states = layer_outputs[0]
            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
            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 = False,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0.0).any():
                return attention_mask
            return None

        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)
            if isinstance(attention_mask, BlockMask):
                return 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)

        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
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        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,
            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
        ):
            min_dtype = torch.finfo(dtype).min
            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,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        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 place the 4D attention mask on.
            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:
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            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()  
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class KwargsForCausalLM(FlashAttentionKwargs): ...


class YoutuForCausalLM(YoutuPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

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

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

        self.post_init()

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

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

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model
    
    def get_merge_embedding(self, inputs_embeds, image_embeds, image_mask,**kwargs,):
        bs, length, dim_size = inputs_embeds.shape
        if image_embeds is None:
            return inputs_embeds
        if bs == 1:
            image_embeds = image_embeds.unsqueeze(0)
            init_inputs_embeds = inputs_embeds.clone()
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
            cmp_mask = torch.isclose(init_inputs_embeds, inputs_embeds, rtol=1e-05, atol=1e-08)
        else:
            assert(bs==1)
        
        return inputs_embeds
    @can_return_tuple
    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
    @add_start_docstrings_to_model_forward(YOUTU_VL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = 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,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> CausalLMOutputWithPast:
        r"""
        Returns:

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

        outputs: BaseModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        
        loss = None

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

class VLPatchMerger(nn.Module):
    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        self.ln_q = YoutuRMSNorm(context_dim, eps=1e-06) 
        self.mlp = nn.Sequential(
            nn.Linear(self.hidden_size, self.hidden_size),
            nn.GELU(),
            nn.Linear(self.hidden_size, dim),
        )
        
    def forward(self, x: torch.Tensor, spatial_shapes: torch.Tensor) -> torch.Tensor:
        x = self.ln_q(x).view(-1, self.hidden_size)
        x = self.mlp(x)
        return x

class YoutuVLForConditionalGeneration(YoutuPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

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

        config.vision_config.out_hidden_size = config.hidden_size
        config.vision_config.vision_use_head = False
        self.siglip2 = Siglip2VisionModel._from_config(config.vision_config)
        self.merger = VLPatchMerger(
            dim=config.hidden_size,
            context_dim=config.vision_config.hidden_size,
            spatial_merge_size=2,
        )
        self.rope_deltas = None

        self.model = YoutuModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.first_logits = None

        self.post_init()


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

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

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model
    
    def get_input_idx_embeddings(self, input_ids):
        inputs_embeds = self.model.embed_tokens(input_ids)
        return inputs_embeds
        
    def get_visiual_features(self, pixel_values, pixel_attention_mask, spatial_shapes):
        pixel_values = pixel_values.type(self.siglip2.dtype)

        # Extract image embeddings via vision model
        image_embeds = self.siglip2(pixel_values, pixel_attention_mask, spatial_shapes).last_hidden_state
        # Merge image features with the output of vision model
        image_embeds = self.merger(image_embeds, spatial_shapes)

        return image_embeds
    
    
    def get_merge_embedding(self, inputs_embeds, image_embeds, image_mask, **kwargs):
        """
        Merge text embeddings with image embeddings using the provided mask.
        
        Args:
            inputs_embeds: Text input embeddings
            image_embeds: Image embeddings to merge
            image_mask: Mask indicating where to place image embeddings
            **kwargs: Additional keyword arguments
            
        Returns:
            Merged embeddings with image features integrated
        """
        bs, length, dim_size = inputs_embeds.shape
        if image_embeds is None:
            return inputs_embeds
        if bs == 1:
            image_embeds = image_embeds.unsqueeze(0)
            init_inputs_embeds = inputs_embeds.clone()
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
            cmp_mask = torch.isclose(init_inputs_embeds, inputs_embeds, rtol=1e-05, atol=1e-08)
        else:
            print('******************ERROR: if you see this info, only support batch_size==1*********************')
            assert(bs == 1)
        
        return inputs_embeds

    @can_return_tuple
    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
    @add_start_docstrings_to_model_forward(YOUTU_VL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = 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,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_attention_mask: Optional[torch.LongTensor] = None,
        spatial_shapes: Optional[torch.LongTensor] = None,
        instance_length: Optional[torch.LongTensor] = None,
        coefficients: Optional[torch.FloatTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:
            TODO: Add example

        Returns:
        """

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

        if inputs_embeds is None:
            inputs_embeds = self.model.embed_tokens(input_ids)
            
            if pixel_values is not None:
                bs, length, dim_size = inputs_embeds.shape
                pixel_values = pixel_values.type(self.siglip2.dtype)
            
                image_embeds = self.siglip2(pixel_values, pixel_attention_mask, spatial_shapes).last_hidden_state
                image_embeds = self.merger(image_embeds, spatial_shapes)

                n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
                n_image_features = image_embeds.shape[0]

                if n_image_tokens > n_image_features:
                    raise ValueError(
                        "Image features and image tokens do not match: tokens: {}, features {}".format(
                            n_image_tokens, n_image_features
                        )
                    )

                mask = input_ids == self.config.image_token_id
                mask_unsqueezed = mask.unsqueeze(-1)
                mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
                image_mask = mask_expanded.to(inputs_embeds.device)
                image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)

                if bs != 1:
                    raise ValueError("Only support batch size = 1")

                image_embeds = image_embeds.unsqueeze(0)
                inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
                                 
            if attention_mask is not None:
                attention_mask = attention_mask.to(inputs_embeds.device)

        outputs: BaseModelOutputWithPast = self.model(
            input_ids=None,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            cache_position=cache_position,
            instance_length=instance_length,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        logits = self.lm_head(hidden_states)
        if logits.shape[1] != 1:
            self.first_logits = logits
        loss = None

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
    def truncate_past_key_values(
        self, 
        past_key_values: Optional[DynamicCache], 
        num_history: int
    ) -> Optional[DynamicCache]:
        """Truncate past_key_values to specified history length in-place.
        
        Args:
            past_key_values: Cache object to truncate
            num_history: Target history length to keep
            
        Returns:
            Truncated cache object or None if input is None
        """
        if past_key_values is None:
            return None
        
        current_length = past_key_values.get_seq_length()
        if current_length <= num_history:
            return past_key_values
        
        for layer_idx in range(len(past_key_values.key_cache)):
            if past_key_values.key_cache[layer_idx] is not None:
                past_key_values.key_cache[layer_idx] = (
                    past_key_values.key_cache[layer_idx][:, :, :num_history, :].contiguous()
                )
                past_key_values.value_cache[layer_idx] = (
                    past_key_values.value_cache[layer_idx][:, :, :num_history, :].contiguous()
                )
        
        return past_key_values
    
    def clone_past_key_values(
        self, 
        past_key_values: Optional[DynamicCache]
    ) -> Optional[DynamicCache]:
        """Deep copy past_key_values to avoid shared reference issues.
        
        Args:
            past_key_values: Cache object to clone
            
        Returns:
            Deep copied cache object or None if input is None
        """
        if past_key_values is None:
            return None
        
        new_cache = DynamicCache()
        for layer_idx in range(len(past_key_values.key_cache)):
            if past_key_values.key_cache[layer_idx] is not None:
                new_cache.key_cache.append(past_key_values.key_cache[layer_idx].clone())
                new_cache.value_cache.append(past_key_values.value_cache[layer_idx].clone())
        
        return new_cache
    
    def concat_token_ids(
        self, 
        input_ids: torch.Tensor, 
        concat_ids: Optional[List[int]]
    ) -> torch.Tensor:
        """Concatenate additional token IDs to input sequence.
        
        Args:
            input_ids: Original input token IDs of shape (batch_size, seq_len)
            concat_ids: Token IDs to concatenate
            
        Returns:
            Concatenated token IDs tensor
        """
        if concat_ids is None:
            return input_ids
        
        num_gen = len(concat_ids)
        if num_gen < 2:
            return input_ids
        
        batch_size = input_ids.size(0)
        concat_token_tensor = torch.tensor(
            concat_ids, 
            dtype=input_ids.dtype, 
            device=input_ids.device
        )
        concat_tokens = concat_token_tensor.unsqueeze(0).repeat(batch_size, 1)
        new_input_ids = torch.cat([input_ids, concat_tokens], dim=1)
        
        return new_input_ids
    
    def create_causal_mask_for_kv_cache(
        self, 
        kv_cache_len: int, 
        num_new_tokens: int, 
        device: torch.device, 
        dtype: torch.dtype = torch.bfloat16
    ) -> torch.Tensor:
        """Create causal attention mask for KV cache usage.
        
        Each new token can only see:
        1. All content in KV cache (positions 0 to kv_cache_len-1)
        2. Previous new tokens and itself (causal masking)
        
        Args:
            kv_cache_len: Length of existing sequence in KV cache
            num_new_tokens: Number of new tokens being added
            device: Target device for tensor allocation
            dtype: Data type for the mask tensor
            
        Returns:
            Attention mask of shape (1, 1, num_new_tokens, kv_cache_len + num_new_tokens)
        """
        total_len = kv_cache_len + num_new_tokens
        min_val = torch.finfo(dtype).min
        
        # Initialize mask with min_val (masked positions)
        mask = torch.full((num_new_tokens, total_len), min_val, device=device, dtype=dtype)
        
        # Set visible positions to 0
        for i in range(num_new_tokens):
            if kv_cache_len > 0:
                mask[i, :kv_cache_len] = 0
            mask[i, kv_cache_len:kv_cache_len + i + 1] = 0
        
        return mask.unsqueeze(0).unsqueeze(0)
    
    def create_4d_causal_mask(
        self, 
        seq_len: int, 
        device: torch.device, 
        dtype: torch.dtype = torch.bfloat16
    ) -> torch.Tensor:
        """Create complete 4D causal attention mask for initial decoding.
        
        Args:
            seq_len: Sequence length
            device: Target device for tensor allocation
            dtype: Data type for the mask tensor
            
        Returns:
            Causal attention mask of shape (1, 1, seq_len, seq_len)
        """
        min_val = torch.finfo(dtype).min
        
        # Create lower triangular causal mask
        mask = torch.full((seq_len, seq_len), min_val, device=device, dtype=dtype)
        mask = torch.triu(mask, diagonal=1)
        
        return mask.unsqueeze(0).unsqueeze(0)
    
    def _first_decoder(
        self, 
        new_input_ids: torch.Tensor, 
        past_key_values: Optional[DynamicCache] = None, 
        image_embeds: Optional[torch.Tensor] = None, 
        image_mask: Optional[torch.Tensor] = None, 
        num_gen: int = 32
    ) -> Tuple[torch.Tensor, Any]:
        """Execute decoder pass with causal attention masking.
        
        This method performs a single decoder pass with optimized attention masking.
        On the first decoding step (when past_key_values is None), it processes image
        embeddings and merges them with text embeddings.
        
        Args:
            new_input_ids: Input token IDs of shape (batch_size, seq_len)
            past_key_values: Cached key-value pairs from previous decoding steps
            image_embeds: Image embeddings to merge (only used in first step)
            image_mask: Mask indicating positions for image embedding placement
            num_gen: Number of tokens to generate in parallel
            
        Returns:
            Tuple containing:
                - predicted_token_ids: Predicted token IDs of shape (batch_size, num_gen)
                - outputs: Model outputs including logits and updated cache
        """
        # Get current sequence position
        start_position = past_key_values.get_seq_length() if past_key_values is not None else 0
        batch_size, seq_len = new_input_ids.shape
        
        # Create position IDs directly on GPU to avoid CPU-GPU transfer
        position_ids = torch.arange(
            start_position, 
            start_position + seq_len,
            dtype=torch.long, 
            device=new_input_ids.device
        ).unsqueeze(0)
        
        # Process image embeddings only on first decoding step
        inputs_embeds = None
        if start_position == 0:
            inputs_embeds = self.get_input_idx_embeddings(new_input_ids)
            if image_embeds is not None:
                inputs_embeds = self.get_merge_embedding(inputs_embeds, image_embeds, image_mask)
        
        # Create 4D causal attention mask
        attention_mask = None
        if start_position > 0 and seq_len > 0:
            # When using KV cache, create mask for new tokens
            attention_mask = self.create_causal_mask_for_kv_cache(
                start_position, seq_len, new_input_ids.device, dtype=torch.bfloat16
            )
        elif start_position == 0 and seq_len > 0:
            # First decoding, create complete causal mask
            attention_mask = self.create_4d_causal_mask(
                seq_len, new_input_ids.device, dtype=torch.bfloat16
            )
        
        with torch.no_grad():
            if start_position > 0:
                outputs = self.forward(
                    input_ids=new_input_ids,
                    inputs_embeds=None,
                    attention_mask=None,  # Note: attention_mask currently disabled
                    position_ids=position_ids,
                    use_cache=True,
                    cache_position=True,
                    past_key_values=past_key_values,
                )
            else:
                outputs = self.forward(
                    input_ids=None,
                    inputs_embeds=inputs_embeds,
                    attention_mask=None,  # Note: attention_mask currently disabled
                    position_ids=position_ids,
                    use_cache=True,
                    cache_position=True,
                    past_key_values=past_key_values,
                )
            
            # Extract predicted token IDs from logits
            predicted_token_ids = outputs.logits[:, -(num_gen + 1):-1].argmax(dim=-1)
        
        return predicted_token_ids, outputs
    
    def generate_parallel_decoder(
        self, 
        inputs: Dict[str, torch.Tensor], 
        image_embeds: torch.Tensor, 
        mask_token_id: int, 
        max_new_tokens: int = 8192, 
        num_gen: int = 64, 
        verbose: bool = False
    ) -> List[int]:
        """Generate tokens using optimized parallel decoding with dual-pass verification.
        
        This method implements a parallel decoding strategy that generates multiple tokens
        simultaneously and verifies them in a second pass. The algorithm:
        1. First pass: Predict tokens with mask tokens
        2. Second pass: Verify predictions with actual predicted tokens
        3. Accept verified tokens and continue from the first unverified position
        
        Optimizations:
        - First decoding uses cloned cache to avoid modifying the original
        - Second decoding updates the original cache in-place
        - Minimizes CPU-GPU data transfers by operating on GPU
        - Pre-allocates tensors to avoid repeated creation
        - Removes debug output from inner loops (controlled by verbose flag)
        - Entire loop wrapped with torch.no_grad() for efficiency
        
        Args:
            inputs: Input dictionary containing 'input_ids' tensor
            image_embeds: Image embeddings for multimodal processing
            mask_token_id: Token ID used for masked positions
            max_new_tokens: Maximum number of tokens to generate
            num_gen: Number of tokens to generate in parallel per iteration
            verbose: If True, print detailed progress information
            
        Returns:
            List of generated token IDs
        """
        if verbose:
            print("Starting parallel decoder generation")
        
        # Constants
        STOP_TOKEN_ID = 128001
        device = self.model.device
        input_ids = inputs["input_ids"]
        decoder_idx = []
        
        # Pre-allocate mask tokens tensor to avoid repeated creation
        mask_tokens = torch.full((1, num_gen), mask_token_id, dtype=torch.long, device=device)
        
        # Initialize KV cache
        prefix_past_key_values = DynamicCache()
        step = 0
        is_exit = False
        
        # Cache initial token ID
        prefix_step_id = input_ids[0, 0].item()
        
        with torch.no_grad():
            while len(decoder_idx) < max_new_tokens and not is_exit:
                # ============ First Pass: Predict with mask tokens ============
                new_input_ids = torch.cat([input_ids, mask_tokens], dim=1)
                
                # Use cloned cache for first pass to preserve original
                if step == 0:
                    first_cache = DynamicCache()
                    
                    # Create image mask for first step
                    mask = new_input_ids == self.config.image_token_id
                    mask_unsqueezed = mask.unsqueeze(-1)
                    mask_expanded = mask_unsqueezed.expand(-1, -1, image_embeds.size(-1))
                    image_mask = mask_expanded.to(image_embeds.device)
                else:
                    first_cache = self.clone_past_key_values(prefix_past_key_values)
                
                first_predicted_ids, _ = self._first_decoder(
                    new_input_ids,
                    past_key_values=first_cache,
                    image_embeds=image_embeds if step == 0 else None,
                    image_mask=image_mask if step == 0 else None,
                    num_gen=num_gen
                )
                
                # ============ Second Pass: Verify with predicted tokens ============
                new_input_ids = torch.cat([input_ids, first_predicted_ids], dim=1)
                
                # Use original cache for second pass (will be updated and retained)
                if step == 0:
                    second_cache = DynamicCache()
                else:
                    second_cache = prefix_past_key_values
                
                second_predicted_ids, outputs = self._first_decoder(
                    new_input_ids,
                    past_key_values=second_cache,
                    image_embeds=image_embeds if step == 0 else None,
                    image_mask=image_mask if step == 0 else None,
                    num_gen=num_gen
                )
                
                # ============ Compare predictions and count successes ============
                first_pred_list = first_predicted_ids[0].tolist()
                second_pred_list = second_predicted_ids[0].tolist()
                
                if verbose:
                    print(f"First pass predictions:  {first_pred_list}")
                    print(f"Second pass predictions: {second_pred_list}")
                
                # Compare predictions to find verified tokens
                success = 0
                for idx in range(len(second_pred_list) - 1):
                    first_id = first_pred_list[idx]
                    second_id = second_pred_list[idx]
                    next_second_id = second_pred_list[idx + 1]
                    
                    # Check for stop token
                    if second_id == STOP_TOKEN_ID:
                        is_exit = True
                        break
                    
                    if next_second_id == STOP_TOKEN_ID and idx == len(second_pred_list) - 2:
                        success += 1
                        is_exit = True
                        break
                    
                    # Verify prediction consistency
                    if first_id == second_id:
                        success += 1
                    else:
                        break
                
                # ============ Update decoded tokens ============
                if step == 0:
                    decoder_idx.extend(second_pred_list[:success])
                else:
                    if verbose:
                        print(f"Verified {success} tokens: {second_pred_list[:success]}")
                    decoder_idx.append(prefix_step_id)
                    decoder_idx.extend(second_pred_list[:success])
                
                if verbose:
                    print(f"Exit status: {is_exit}")
                    print(f"Total decoded tokens: {len(decoder_idx)}")
                
                # ============ Truncate KV cache to verified length ============
                past_key_values = outputs.past_key_values
                if past_key_values is not None:
                    current_kv_len = past_key_values.get_seq_length()
                    num_to_keep = current_kv_len - (num_gen - success)
                    prefix_past_key_values = self.truncate_past_key_values(
                        past_key_values, num_to_keep
                    )
                else:
                    prefix_past_key_values = None
                
                # Update input_ids for next iteration
                next_token_id = (
                    second_pred_list[success] 
                    if success < len(second_pred_list) 
                    else prefix_step_id
                )
                input_ids = torch.tensor(
                    [[next_token_id]], 
                    dtype=torch.long, 
                    device=device
                )
                prefix_step_id = next_token_id
                
                step += 1
                
                if verbose:
                    print(f"Step {step} completed, success rate: {success}/{num_gen}\n")
        
        return decoder_idx

__all__ = ["YoutuPreTrainedModel", "YoutuModel", "YoutuVLForConditionalGeneration"]