modeling_sdar.py

# This file is modified based on https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen3/modeling_qwen3.py.
#
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/qwen3/modular_qwen3.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_qwen3.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
#
# 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.

from typing import Callable, Optional, Tuple, Union, List

import torch
from torch import nn
from einops import rearrange

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
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 LossKwargs, auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
from configuration_sdar import SDARConfig
from fused_linear_diffusion_cross_entropy import FusedLinearDiffusionCrossEntropyLoss

from flash_attn.ops.triton.layer_norm import rms_norm_fn as flash_rms_norm

import torch.nn.functional as F
try:
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
except:
    pass

try:
    from liger_kernel.ops.swiglu import LigerSiLUMulFunction  # noqa: F401
    liger_kernel_is_available = True
except ImportError:
    liger_kernel_is_available = False


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


logger = logging.get_logger(__name__)


def modify_padded_position_ids_2d(position_ids: torch.LongTensor) -> torch.LongTensor:
    """
    使用完全向量化的 PyTorch 操作修改一个 batch 的 packed position_ids。
    这个函数假设输入是一个 2D Tensor，形状为 (batch_size, sequence_length)。
    它会独立地处理 batch 中的每一行。

    Args:
        position_ids: 二维 PyTorch Tensor, shape (batch_size, sequence_length).

    Returns:
        修改后的 position_ids Tensor, shape (batch_size, sequence_length).
    """
    if position_ids.dim() != 2:
        raise ValueError(f"Input tensor must be 2D, but got {position_ids.dim()} dimensions.")
        
    batch_size, seq_len = position_ids.shape
    device = position_ids.device

    col_indices = torch.arange(seq_len, device=device, dtype=position_ids.dtype).expand(batch_size, -1)
    mask = (position_ids != 0)

    masked_indices = col_indices * mask
    last_nonzero_idx = torch.max(masked_indices, dim=1).values
    has_nonzero = torch.any(mask, dim=1)
    pad_start_idx = torch.where(has_nonzero, last_nonzero_idx + 1, torch.tensor(0, device=device, dtype=position_ids.dtype))

    padding_mask = col_indices >= pad_start_idx.unsqueeze(1)
    new_pad_values = col_indices - pad_start_idx.unsqueeze(1)
    position_ids = torch.where(padding_mask, new_pad_values, position_ids)

    return position_ids


def calculate_token_nums(position_ids: torch.Tensor):
    """
    使用 PyTorch 高效计算一个批次中每个打包序列的长度。

    Args:
        position_ids (torch.Tensor): 一个 2D Tensor，形状为 (batch_size, sequence_length)。
                                     例如：tensor([[0,1,2,3,4,0,1,2,3,4,5,0,1,2,3,0,0,0]])
    Returns:
        list[list[int]]: 一个嵌套列表，包含每个批次项中各个序列的长度。
                         例如：[[5, 6, 4, 1, 1, 1]]
    """
    # 检查输入是否为 2D Tensor
    if position_ids.dim() != 2:
        raise ValueError(f"输入必须是 2D Tensor，但得到了 {position_ids.dim()}D")

    all_lengths = []
    
    # 我们按批次逐行处理。因为每行的序列长度数量不同（ragged），
    # 所以 Python 循环在批次维度上是最高效且最清晰的写法。
    # 循环内部的操作是完全向量化的。
    for pids_row in position_ids:
        # 获取当前行的总长度
        seq_len = pids_row.shape[0]
        
        # 1. 找到所有值为 0 的元素的索引
        # pids_row == 0 会返回一个布尔 Tensor: [True, False, ..., True, ...]
        # torch.nonzero 会返回这些 True 值的索引
        # .flatten() 将其从 (N, 1) 形状的 Tensor 变为 (N,) 形状
        zero_indices = torch.nonzero(pids_row == 0).flatten()
        
        # 2. 将序列的总长度作为一个额外的切分点添加到末尾
        # 这对于计算最后一个序列的长度至关重要
        # 注意：要确保新创建的 tensor 和原始 tensor 在同一个设备上 (cpu/cuda)
        split_points = torch.cat([
            zero_indices,
            torch.tensor([seq_len], device=pids_row.device, dtype=zero_indices.dtype)
        ])
        
        # 3. 计算相邻切分点之间的差值，这就是我们想要的长度
        # torch.diff([a, b, c, d]) 会返回 [b-a, c-b, d-c]
        lengths = torch.diff(split_points)

        all_lengths.append(lengths)

    return all_lengths


def forward_add_noise_packed(
    inputs_ids: torch.Tensor,
    num_tokens_list: List[torch.Tensor],
    prompt_mask: torch.Tensor,
    mask_id: int,
    eps: float = 1e-3,
    max_tries: int = 10,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    为一批打包（packed）序列的 token ID 添加噪声。

    此函数保留了为每个逻辑样本（在每个批次项内拼接）生成独立随机噪声率的逻辑。
    它会随机将一部分 token 的 ID 替换为 mask_id。
    这个过程会避开被 prompt_mask 标记的位置。

    Args:
        inputs_ids (torch.Tensor): 
            输入的 token ID 张量，形状为 (bsz, total_tokens)。
        num_tokens_list (List[torch.Tensor]): 
            一个张量列表，长度为 bsz。列表中的每个张量记录了对应批次项中
            每个逻辑样本的长度。例如: [tensor([len1, len2]), tensor([len3, len4, len5])].
        prompt_mask (torch.Tensor): 
            布尔型张量，形状为 (bsz, total_tokens)，值为 True 的位置表示是 prompt，
            不应添加噪声。
        mask_id (int): 
            用于替换的 mask token 的 ID。
        eps (float): 
            微小值，用于防止噪声率 t 恰好为 0，确保 p_mask > 0。
        max_tries (int): 
            为确保至少一个非 prompt token 被 mask，对每个批次项尝试的最大次数。

    Returns:
        Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        - noisy_input_ids (torch.Tensor): 
            添加噪声后的 token ID 张量，形状为 (bsz, total_tokens)。
        - final_masked_indices (torch.Tensor): 
            布尔型张量，标记了哪些位置被实际 mask 了，形状为 (bsz, total_tokens)。
        - p_masks (torch.Tensor): 
            一个一维张量，包含了被 mask 的 token 对应的实际噪声率。
    """
    # 1. 验证和获取形状
    bsz, total_tokens = inputs_ids.shape
    device = inputs_ids.device

    # 检查输入的一致性
    assert len(num_tokens_list) == bsz, f"num_tokens_list 的长度 ({len(num_tokens_list)}) 必须等于 bsz ({bsz})"
    assert prompt_mask.shape == (bsz, total_tokens), f"prompt_mask 形状不匹配, 期望 {(bsz, total_tokens)}, 得到 {prompt_mask.shape}"

    # 准备结果容器
    noisy_ids_list = []
    final_masked_indices_list = []
    p_masks_per_token_list = []

    # 2. 在批次维度上迭代
    # 这是处理不同打包结构最直接有效的方法
    for i in range(bsz):
        # 提取当前批次项的数据
        current_ids = inputs_ids[i:i+1] # shape: (1, total_tokens)
        current_num_tokens = num_tokens_list[i]
        current_prompt_mask = prompt_mask[i:i+1] # shape: (1, total_tokens)
        
        num_samples_in_item = len(current_num_tokens)
        # 验证当前批次项的 token 总数是否匹配
        assert total_tokens == torch.sum(current_num_tokens), \
            f"批次项 {i} 的 num_tokens 之和 ({torch.sum(current_num_tokens)}) 与 total_tokens ({total_tokens}) 不匹配"

        eligible_for_masking = ~current_prompt_mask

        # 如果没有任何 token 可以被 mask，直接使用原始输入，并设置 p_mask 为 eps
        if not eligible_for_masking.any():
            noisy_ids_list.append(current_ids)
            final_masked_indices_list.append(torch.zeros_like(current_prompt_mask, dtype=torch.bool))
            # p_mask_per_token 的形状应为 (1, total_tokens) 以便后续拼接
            p_masks_per_token_list.append(torch.full((1, total_tokens), eps, device=device, dtype=torch.float))
            continue

        # --- 尝试生成 mask，确保至少 mask 一个 token ---
        final_masked_indices_item = torch.zeros_like(current_prompt_mask, dtype=torch.bool)
        p_mask_per_token = None
        
        for _ in range(max_tries):
            # 为每个逻辑样本生成一个独立的噪声率 t
            t = torch.rand(num_samples_in_item, device=device)
            p_mask_per_sample = (1 - eps) * t + eps

            # 将每个样本的噪声率扩展到其所有 token 上
            p_mask_per_token_1d = torch.repeat_interleave(p_mask_per_sample, current_num_tokens)
            p_mask_per_token = p_mask_per_token_1d.unsqueeze(0) # shape: (1, total_tokens)

            # 根据噪声率生成随机 mask
            masked_indices = torch.rand_like(p_mask_per_token) < p_mask_per_token
            # 应用 prompt mask，确保 prompt 不被 mask
            final_masked_indices_item = masked_indices & eligible_for_masking

            # 如果成功 mask 了至少一个 token，则跳出尝试循环
            if final_masked_indices_item.any():
                break
        
        # 如果 max_tries 之后仍然没有 mask 任何 token (极小概率)，就强制 mask 一个可 mask 的 token
        if not final_masked_indices_item.any():
            eligible_indices = torch.nonzero(eligible_for_masking.squeeze(0), as_tuple=True)[0]
            if len(eligible_indices) > 0:
                # 随机选择一个可 mask 的位置
                random_choice = torch.randint(0, len(eligible_indices), (1,)).item()
                force_mask_idx = eligible_indices[random_choice]
                final_masked_indices_item[0, force_mask_idx] = True


        # --- 根据最终的 mask 生成带噪声的 IDs ---
        noisy_ids_item = torch.where(
            final_masked_indices_item,
            mask_id,
            current_ids
        )
        
        # 保存这个批次项的结果
        noisy_ids_list.append(noisy_ids_item)
        final_masked_indices_list.append(final_masked_indices_item)
        p_masks_per_token_list.append(p_mask_per_token)

    # 3. 将列表中的结果堆叠成最终的批处理张量
    noisy_input_ids = torch.cat(noisy_ids_list, dim=0)
    final_masked_indices = torch.cat(final_masked_indices_list, dim=0)
    p_mask_full = torch.cat(p_masks_per_token_list, dim=0)
    
    # 4. 提取被 mask 位置对应的噪声率
    p_masks = p_mask_full[final_masked_indices]
    
    return noisy_input_ids, final_masked_indices, p_masks


def block_diff_mask(b, h, q_idx, kv_idx, block_size=None, n=None):
    """
    Constructs the specialized block diffusion attention mask for training
    composed of three masks:
    - **Block Diagonal Mask (M_BD)**: Self-attention within noised blocks
    - **Offset Block Causal Mask (M_OBC)**: Cross-attention for conditional context
    - **Block Causal Mask (M_BC)**: Attention to update x0

    Args:
        b, h: Batch and head indices (ignored for mask logic).
        q_idx, kv_idx: Query and Key indices.
        seq_len: Total sequence length.
        block_size: Defines the block structure.

    Returns:
        A boolean attention mask.
    """

    # Indicate whether token belongs to xt or x0
    x0_flag_q = q_idx >= n
    x0_flag_kv = kv_idx >= n

    # Compute block indices
    block_q = torch.where(
        x0_flag_q == 1, (q_idx - n) // block_size, q_idx // block_size
    )
    block_kv = torch.where(
        x0_flag_kv == 1, (kv_idx - n) // block_size, kv_idx // block_size
    )

    # **1. Block Diagonal Mask (M_BD) **
    block_diagonal = (block_q == block_kv) & (x0_flag_q == x0_flag_kv)

    # **2. Offset Block-Causal Mask (M_OBC) **
    offset_block_causal = (block_q > block_kv) & (
        x0_flag_kv == 1) & (x0_flag_q == 0)

    # **3. Block-Causal Mask (M_BC) **
    block_causal = (block_q >= block_kv) & (x0_flag_kv == 1) & (x0_flag_q == 1)

    # **4. Combine Masks **
    return block_diagonal | offset_block_causal | block_causal


def block_attn_mask(num_tokens, block_size, device):
    masks = []
    for i in range(len(num_tokens)):
        cur_masks = []
        for num in num_tokens[i]:
            # 全部返回 n*n 而非 2n*2n
            single_mask = block_diff_mask(
                b=None,
                h=None,
                q_idx=torch.arange(num * 2, device=device)[:, None],
                kv_idx=torch.arange(num * 2, device=device)[None, :],
                block_size=block_size,
                n=num,
            )
            cur_masks.append(single_mask)
        masks.append(torch.block_diag(*cur_masks))
    masks = torch.stack(masks, dim=0)
    return masks


@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
def fused_flex_attention(query, key, value, attention_mask, **kwargs):
    return flex_attention(query, key, value, block_mask=attention_mask, **kwargs)


@use_kernel_forward_from_hub("RMSNorm")
class SDARRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        SDARRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        return flash_rms_norm(
            hidden_states, weight=self.weight, bias=None, eps=self.variance_epsilon)
        '''
        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 SDARMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(
            self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(
            self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(
            self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        if liger_kernel_is_available:
            return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
        else:
            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


class SDARAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: SDARConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(
            config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.hidden_size = config.hidden_size
        self.num_attention_heads = config.num_attention_heads
        self.num_key_value_heads = config.num_key_value_heads

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        # unlike olmo, only on the head dim!
        self.q_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        # thus post q_norm does not need reshape
        self.k_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.sliding_window = config.sliding_window
        if not (
            self.config.use_sliding_window
            and getattr(self.config, "sliding_window", None) is not None
            and self.layer_idx >= self.config.max_window_layers
        ):
            self.sliding_window = None

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        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]]]:
        input_shape = hidden_states.shape[:-1]
        bsz, q_len = input_shape
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_norm(self.q_proj(
            hidden_states).view(hidden_shape)).transpose(1, 2)
        key_states = self.k_norm(self.k_proj(
            hidden_states).view(hidden_shape)).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(
            hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(
            query_states, key_states, cos, sin)

        if past_key_value is not None and kwargs.get("store_kv", False):
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            key_states, value_states = past_key_value.update(
                key_states, value_states, self.layer_idx)
        elif past_key_value is not None and not kwargs.get("store_kv", False) and len(past_key_value) > self.layer_idx:
            # only retrive, do not store kv
            past_key_states, past_value_states = past_key_value[self.layer_idx]
            key_states = torch.cat(
                [past_key_states, key_states], dim=-2)
            value_states = torch.cat(
                [past_value_states, value_states], dim=-2)

        if self.training:
            attn_output, attn_weights = fused_flex_attention(
                query=query_states,
                key=key_states,
                value=value_states,
                attention_mask=attention_mask,
                enable_gqa=True,
                scale=self.scaling,
                return_lse=True
            )
            attn_weights = attn_weights.to(
                value_states.dtype) if attn_weights is not None else None
            attn_output = rearrange(attn_output, 'b h l d -> b l (h d)')
        else:
            attention_mask = attention_mask.bool() if attention_mask is not None else None
            attn_weights = None
            if torch.all(attention_mask):  # decoding
                query_states = query_states.transpose(1, 2)
                key_states = key_states.transpose(1, 2)
                value_states = value_states.transpose(1, 2)
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    causal=False,
                    softmax_scale=self.scaling
                )
                attn_output = rearrange(attn_output, 'b l h d -> b l (h d)')
            else:  # prefilling
                attn_output = F.scaled_dot_product_attention(
                    query=query_states,
                    key=key_states,
                    value=value_states,
                    attn_mask=attention_mask,
                    is_causal=False,
                    scale=self.scaling,
                    enable_gqa=True
                )
                attn_output = rearrange(attn_output, 'b h l d -> b l (h d)')
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights  # , attn_weights


class SDARDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: SDARConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = SDARAttention(config=config, layer_idx=layer_idx)
        self.mlp = SDARMLP(config)
        self.input_layernorm = SDARRMSNorm(
            config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = SDARRMSNorm(
            config.hidden_size, eps=config.rms_norm_eps)
        if (
            config.sliding_window and config._attn_implementation != "flash_attention_2"
        ):  # diff with Llama is this warning
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )

    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,
        use_cache: Optional[bool] = False,
        store_kv: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        # necessary, but kept here for BC
        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)

        # Self Attention
        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,
            use_cache=use_cache,
            store_kv=store_kv,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

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

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


@auto_docstring
class SDARPreTrainedModel(PreTrainedModel):
    config_class = SDARConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["SDARDecoderLayer"]
    _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, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, SDARRMSNorm):
            module.weight.data.fill_(1.0)


class SDARRotaryEmbedding(nn.Module):
    def __init__(self, config: SDARConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        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()
    # power user: used with advanced RoPE types (e.g. dynamic rope)
    @dynamic_rope_update
    def forward(self, x, position_ids):
        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)


@auto_docstring
class SDARModel(SDARPreTrainedModel):
    def __init__(self, config: SDARConfig):
        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(
            [SDARDecoderLayer(config, layer_idx)
             for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = SDARRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = SDARRotaryEmbedding(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
    @auto_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,
        store_kv: Optional[bool] = 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 self.gradient_checkpointing and self.training and use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
            )
            use_cache = False

        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
        if not isinstance(past_key_values, (type(None), Cache)):
            raise ValueError(
                "The `past_key_values` should be either a `Cache` object or `None`.")

        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

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

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

        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=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                store_kv=store_kv,
                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)

        # add hidden states from the last decoder layer
        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: Union[torch.Tensor, "BlockMask"],
        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 past_key_values is not None:
                is_padding_right = attention_mask[:, -
                                                  1].sum().item() != input_tensor.size()[0]
                if is_padding_right:
                    raise ValueError(
                        "You are attempting to perform batched generation with padding_side='right'"
                        " this may lead to unexpected behaviour for Flash Attention version of Qwen3. Make sure to "
                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                    )
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None
        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                seq_len_q, seq_len_kv = attention_mask.shape
                assert seq_len_q == seq_len_kv, f"got {attention_mask.shape=}"
                attention_mask = create_block_mask(
                    # 2d bool tensor, shape: [2*seqlen, 2*seqlen]
                    lambda b, h, q_idx, kv_idx: attention_mask[q_idx, kv_idx],
                    B=None, H=None, Q_LEN=seq_len_q, KV_LEN=seq_len_kv,
                )
            else:
                # Here we pass in flex mask computed externally
                assert isinstance(attention_mask, BlockMask)
            return attention_mask

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

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

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

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

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

        return causal_mask

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

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
            config (`SDARConfig`):
                The model's configuration class
            past_key_values (`Cache`):
                The cache class that is being used currently to generate
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
            )
            diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
                -1, 1
            )
            text_config = config.get_text_config()
            if getattr(text_config, "use_sliding_window", True) and text_config.sliding_window is not None:
                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
                # the check is needed to verify is current checkpoint was trained with sliding window or not
                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
                    sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= (
                        cache_position.reshape(-1, 1) -
                        text_config.sliding_window
                    )
                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
            causal_mask *= diagonal_attend_mask
            causal_mask = causal_mask[None, None,
                                      :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                if attention_mask.shape[-1] > target_length:
                    attention_mask = attention_mask[:, :target_length]
                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, LossKwargs):
    ...


@auto_docstring
class SDARForCausalLM(SDARPreTrainedModel, 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 = SDARModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(
            config.hidden_size, config.vocab_size, bias=False)

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

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

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

    def get_output_embeddings(self):
        return self.lm_head

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

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

    def get_decoder(self):
        return self.model

    def prepare_for_bd_training(self, inputs_ids, position_ids, prompt_mask, masked_indices=None, p_mask_input=None):
        bsz, seq_len = inputs_ids.shape
        num_tokens = calculate_token_nums(position_ids) # List[torch.Tensor]
        
        # 如果手动传入了 masked_indices，就直接用它
        if masked_indices is not None:
            # 手动mask模式：用于RL训练或固定mask实验
            # 注意：外部传入的masked_indices已经只在response部分（通过 & response_mask），不需要再次过滤
            noisy_inputs_ids = torch.where(masked_indices, self.config.mask_token_id, inputs_ids)
            logits_to_keep_half = masked_indices  # (B, L) bool
            # 生成默认的p_mask：扁平化后的噪声率，形状为(M,)，其中M=sum(masked_indices)
            # 默认值0.5表示中等噪声水平（用于扩散loss）
            M = masked_indices.sum().item()
            p_mask = torch.full((M,), 0.5, device=inputs_ids.device, dtype=torch.float)
        else:
            # 随机mask模式：用于Block Diffusion预训练
            # 返回：noisy_inputs_ids (B, L), logits_to_keep_half (B, L) bool, p_mask (M,) float
            noisy_inputs_ids, logits_to_keep_half, p_mask = forward_add_noise_packed(
                inputs_ids=inputs_ids,
                num_tokens_list=num_tokens,
                prompt_mask=prompt_mask,
                mask_id=self.config.mask_token_id,
            )
        
        # 确保两个分支返回的形状一致
        # logits_to_keep_half: (B, L) bool - 标记哪些位置被mask
        # p_mask: (M,) float - 每个被mask位置的噪声率，其中M = sum(logits_to_keep_half)
        assert logits_to_keep_half.shape == (bsz, seq_len), f"logits_to_keep_half shape error: {logits_to_keep_half.shape}"
        assert p_mask.shape == (logits_to_keep_half.sum(),), f"p_mask shape error: {p_mask.shape}, expected ({logits_to_keep_half.sum()},)"
        
        # 如果提供了p_mask_input（用于RL训练），计算p_to_keep
        # p_to_keep表示从masked位置中选出p_mask=True的位置
        p_to_keep = None
        if p_mask_input is not None:
            # 注意：外部传入的p_mask_input已经只在response部分（通过 & response_mask），不需要再次过滤
            # p_mask_input (B, L), logits_to_keep_half (B, L)
            # p_to_keep (M,) bool，其中M=sum(logits_to_keep_half)
            p_to_keep = p_mask_input[logits_to_keep_half]
        
        router_noisy_part_list = []
        for i in range(bsz):
            cur_router_noisy_part = (torch.arange(num_tokens[i].shape[0] *2) % 2 == 0).to(inputs_ids.device)
            cur_router_noisy_part = cur_router_noisy_part.repeat_interleave(num_tokens[i].repeat_interleave(2))
            router_noisy_part_list.append(cur_router_noisy_part)
        router_noisy_part = torch.stack(router_noisy_part_list, dim=0)

        # concated inputs_ids: (bzs, seq_len x 2)
        concat_inputs_ids = inputs_ids.repeat(1, 2)
        # concated logits_to_keep: (bsz, seq_len x 2)
        logits_to_keep = torch.zeros(
                    bsz, 2 * seq_len, dtype=torch.bool, device=inputs_ids.device)
        # concated position_ids: (bsz, seq_len x 2)
        concat_position_ids = torch.zeros(
                    bsz, 2 * seq_len, dtype=position_ids.dtype, device=position_ids.device)
        for i in range(bsz):
            concat_inputs_ids[i][router_noisy_part[i]] = noisy_inputs_ids[i]
            concat_inputs_ids[i][~router_noisy_part[i]] = inputs_ids[i]

            logits_to_keep[i][router_noisy_part[i]] = logits_to_keep_half[i]

            concat_position_ids[i][router_noisy_part[i]] = position_ids[i]
            concat_position_ids[i][~router_noisy_part[i]] = position_ids[i]

        # create flex_attention mask
        attention_mask = block_attn_mask(num_tokens, self.config.block_size, inputs_ids.device)
        flex_attention_mask_3d = create_block_mask(
                            lambda b, h, q_idx, kv_idx: attention_mask[b, q_idx, kv_idx],
                            B=attention_mask.size(0), H=None,
                            Q_LEN=attention_mask.size(1), KV_LEN=attention_mask.size(2),
        )

        return concat_inputs_ids, concat_position_ids, flex_attention_mask_3d, logits_to_keep_half, logits_to_keep, p_mask, p_to_keep

    @can_return_tuple
    @auto_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,
        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,
        masked_indices: Optional[torch.Tensor] = None,
        return_logits: bool = False,
        # RL training parameters
        compute_rl_loss: bool = False,
        p_mask: Optional[torch.Tensor] = None,
        adv: Optional[torch.Tensor] = None,
        adv_optimization: bool = False,
        logp_old_tok: Optional[torch.Tensor] = None,
        logp_ref_tok: Optional[torch.Tensor] = None,
        is_real: Optional[torch.Tensor] = None,
        ppo_eps: float = 0.2,
        kl_beta: float = 0.0,
        use_kl_estimator_k3: bool = True,
        return_entropy: bool = False,
        dynamic_threshold: Optional[float] = None,
        loss_mean: bool = True,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> 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
        )

        if self.training:
            assert inputs_embeds is None, "only support input_ids during training"

            prompt_mask = (labels == -100) if labels is not None else None
            position_ids = modify_padded_position_ids_2d(position_ids)

            (
                concat_inputs_ids,
                concat_position_ids,
                flex_attention_mask_3d,
                logits_to_keep_half,
                logits_to_keep,
                p_mask_out,
                p_to_keep,
            ) = self.prepare_for_bd_training(
                input_ids, position_ids, prompt_mask, masked_indices, p_mask_input=p_mask
            )

            outputs = self.model(
                input_ids=concat_inputs_ids,
                attention_mask=flex_attention_mask_3d,
                position_ids=concat_position_ids,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=True,
                cache_position=cache_position,
                **kwargs,
            )

            hidden_states = outputs.last_hidden_state
            hidden_states = hidden_states[logits_to_keep].contiguous()

            # 初始化 entropy
            entropy = torch.tensor(0.0, device=input_ids.device)

            # ====================== RL loss（PPO） ======================
            if compute_rl_loss:
                assert p_to_keep is not None, "p_mask must be provided for RL loss computation."
                assert adv is not None, "adv must be provided for RL loss computation."
                assert is_real is not None, "is_real must be provided for RL loss computation."
                assert labels is not None, "labels must be provided for RL loss computation."
                assert masked_indices is not None, "masked_indices must be provided for RL loss computation."

                device = input_ids.device

                # logits (M, V) — 保持原样
                logits = self.lm_head(hidden_states)

                # mask — 保持原样
                is_real_tensor = (
                    is_real.to(device=device, dtype=torch.bool)
                    if torch.is_tensor(is_real)
                    else torch.tensor(is_real, dtype=torch.bool, device=device)
                )
                p_mask_real = p_mask & is_real_tensor.unsqueeze(1)      # (B, L)
                p_to_keep_real = p_mask_real[masked_indices]            # (M,) bool

                # 选出 logits — 保持原样
                logits_p = logits[p_to_keep_real]                       # (N, V)
                N = p_to_keep_real.sum().item()
                total_response_tokens = (labels != -100).sum().item()
                total_p_mask = p_mask.sum().item()
                total_masked_indices = masked_indices.sum().item()
                total_is_real = is_real_tensor.sum().item() if is_real_tensor.dim() > 0 else (1 if is_real_tensor.item() else 0)


                # log_softmax
                log_probs_p = torch.nn.functional.log_softmax(logits_p, dim=-1)

                # labels / logp — 保持原样
                labels_p = labels[masked_indices][p_to_keep_real]        # (N,)
                logp_p = log_probs_p.gather(dim=-1, index=labels_p.unsqueeze(-1)).squeeze(-1)

                # entropy（可选）
                if return_entropy:
                    with torch.no_grad():
                        entropy_p = -(log_probs_p.exp() * log_probs_p).sum(dim=-1)
                        entropy = entropy_p.mean() if entropy_p.numel() > 0 else torch.tensor(0.0, device=device)
                        del entropy_p

                # advantage 处理
                adv_tensor = adv.to(device) if torch.is_tensor(adv) else torch.tensor(adv, dtype=torch.float, device=device)
                adv_optimization=False
                if adv_optimization:
                    # token级别优化：对相同前缀取最大advantage（剪枝优化版本）
                    response_mask = (labels != -100)  # (B, L)
                    bsz, seq_len = input_ids.shape
                    
                    # 预计算每个样本的response起始位置
                    response_starts = torch.full((bsz,), seq_len, dtype=torch.long, device=device)
                    for b in range(bsz):
                        if response_mask[b].any():
                            response_starts[b] = response_mask[b].long().argmax()
                    
                    # 剪枝1: 找出已经是最大advantage的样本，直接填充不参与比较
                    max_adv_value = adv_tensor.max()
                    is_max_adv = (adv_tensor == max_adv_value)  # (B,) bool
                    
                    # 创建优化后的 advantage map (B, L)，确保dtype与adv_tensor一致
                    optimized_adv = torch.zeros_like(labels, dtype=adv_tensor.dtype)
                    
                    # 对于已是最大advantage的样本，直接填充
                    for b in range(bsz):
                        if is_max_adv[b]:
                            optimized_adv[b][response_mask[b]] = max_adv_value
                    
                    # 统计信息
                    total_response_tokens = 0
                    updated_tokens = 0
                    skipped_tokens = 0
                    original_adv_sum = 0.0
                    optimized_adv_sum = 0.0
                    
                    # 按position处理，批量比较前缀
                    for pos in range(seq_len):
                        valid_samples = response_mask[:, pos]  # (B,)
                        if not valid_samples.any():
                            continue
                        
                        # 剪枝2: 排除已是最大advantage的样本
                        valid_samples = valid_samples & ~is_max_adv
                        if not valid_samples.any():
                            # 所有样本都是最大值，统计后跳过
                            max_count = (response_mask[:, pos] & is_max_adv).sum().item()
                            total_response_tokens += max_count
                            skipped_tokens += max_count
                            original_adv_sum += max_adv_value.item() * max_count
                            optimized_adv_sum += max_adv_value.item() * max_count
                            continue
                        
                        # 获取所有需要处理的样本索引
                        valid_indices = valid_samples.nonzero(as_tuple=True)[0]  # (N,)
                        
                        for b in valid_indices:
                            b_item = b.item()
                            response_start = response_starts[b_item].item()
                            prefix_len = pos + 1 - response_start
                            
                            if prefix_len <= 0:
                                optimized_adv[b_item, pos] = adv_tensor[b_item]
                                continue
                            
                            # 找出所有response起始位置相同且在pos位置有效的样本（包括已是最大值的）
                            same_start_mask = (response_starts == response_start) & response_mask[:, pos]
                            same_start_indices = same_start_mask.nonzero(as_tuple=True)[0]
                            
                            if len(same_start_indices) == 1:
                                # 只有自己，不需要比较
                                optimized_adv[b_item, pos] = adv_tensor[b_item]
                                total_response_tokens += 1
                                original_adv_sum += adv_tensor[b_item].item()
                                optimized_adv_sum += adv_tensor[b_item].item()
                                continue
                            
                            # 剪枝3: 如果候选中有最大advantage样本，可以直接用最大值
                            has_max_in_candidates = (same_start_mask & is_max_adv).any()
                            
                            prefix_end = pos + 1
                            current_prefix = input_ids[b_item, response_start:prefix_end]
                            
                            # 批量比较：提取所有候选样本的前缀
                            prefixes = input_ids[same_start_indices, response_start:prefix_end]  # (M, prefix_len)
                            
                            # 使用广播比较：(M, prefix_len) vs (prefix_len,)
                            matches = (prefixes == current_prefix.unsqueeze(0)).all(dim=1)  # (M,)
                            
                            # 找到匹配的样本
                            matching_indices = same_start_indices[matches]
                            
                            # 在相同前缀的样本中取最大 advantage
                            original_adv_value = adv_tensor[b_item].item()
                            if matching_indices.numel() > 0:
                                # 剪枝4: 如果匹配中有最大值样本，直接用最大值
                                if has_max_in_candidates and is_max_adv[matching_indices].any():
                                    max_adv = max_adv_value
                                else:
                                    max_adv = adv_tensor[matching_indices].max()
                                
                                optimized_adv[b_item, pos] = max_adv
                                # 统计
                                if abs(max_adv.item() - original_adv_value) > 1e-6:
                                    updated_tokens += 1
                                original_adv_sum += original_adv_value
                                optimized_adv_sum += max_adv.item()
                            else:
                                optimized_adv[b_item, pos] = adv_tensor[b_item]
                                original_adv_sum += original_adv_value
                                optimized_adv_sum += original_adv_value
                            
                            total_response_tokens += 1
                    
                    # 输出统计信息
                    if total_response_tokens > 0:
                        update_ratio = updated_tokens / total_response_tokens
                        skip_ratio = skipped_tokens / total_response_tokens
                        avg_original = original_adv_sum / total_response_tokens
                        avg_optimized = optimized_adv_sum / total_response_tokens
                        print(f"[Adv Optimization] Total: {total_response_tokens}, "
                              f"Updated: {updated_tokens} ({update_ratio:.2%}), "
                              f"Skipped: {skipped_tokens} ({skip_ratio:.2%}), "
                              f"Avg adv: {avg_original:.4f} -> {avg_optimized:.4f} "
                              f"(+{avg_optimized - avg_original:.4f})")
                    
                    # 使用优化后的 advantage
                    adv_expanded = optimized_adv
                else:
                    # 不优化：直接使用原始 advantage
                    adv_expanded = adv_tensor.unsqueeze(1).expand_as(p_mask)
                
                adv_p = adv_expanded[masked_indices][p_to_keep_real]

                # old logp
                if logp_old_tok is not None and logp_old_tok.numel() > 0:
                    logp_old_p = logp_old_tok.to(device)[masked_indices][p_to_keep_real]
                else:
                    logp_old_p = logp_p.detach()

                # ratio/exp
                ratio_p = (logp_p - logp_old_p).clamp(-10.0, 10.0).exp()
                clipped = ratio_p.clamp(1 - ppo_eps, 1 + ppo_eps+0.08)
                surrogate_p = torch.minimum(ratio_p * adv_p, clipped * adv_p)
                # 输出离1最远的ratio值
                # if not torch.allclose(ratio_p, torch.ones_like(ratio_p)):
                furthest_value = ratio_p[torch.abs(ratio_p - 1).argmax()]
                # print(f"Furthest ratio from 1: {furthest_value.item()}")

                # Policy loss: use mean or sum based on loss_mean parameter
                num_masked = masked_indices.sum().item()
                num_loss_elements = surrogate_p.numel()
                print(f"masked_indices.sum()={num_masked}, surrogate_p.numel()={num_loss_elements}")
                if loss_mean:
                    policy_loss = -surrogate_p.mean()
                else:
                    policy_loss = -surrogate_p.sum()

                # KL（可选）
                kl_loss = torch.tensor(0.0, device=device)
                if kl_beta > 0 and logp_ref_tok is not None:
                    logp_ref_p = logp_ref_tok.to(device)[masked_indices][p_to_keep_real]
                    kl_seq_p = logp_p - logp_ref_p

                    if use_kl_estimator_k3:
                        kl_seq_p = (-kl_seq_p).clamp(-10.0, 10.0).exp() - 1.0 + kl_seq_p

                    # KL loss: use mean or sum based on loss_mean parameter
                    if loss_mean:
                        kl_loss = kl_beta * kl_seq_p.mean()
                    else:
                        kl_loss = kl_beta * kl_seq_p.sum()
                    del logp_ref_p, kl_seq_p

                loss = policy_loss + kl_loss
                kl_loss_value = kl_loss.detach().clone()

                # 清理
                del logits, logits_p, log_probs_p, labels_p
                del is_real_tensor, p_mask_real, p_to_keep_real
                del adv_tensor, adv_expanded, adv_p
                del logp_p, logp_old_p, ratio_p, clipped, surrogate_p
                del policy_loss, kl_loss

                logits = None

            # ====================== GRPO / return logits ======================
            elif return_logits:
                logits = self.lm_head(hidden_states)
                loss = None

            # ====================== Block Diffusion fused loss ======================
            else:
                assert labels is not None, "Labels must be provided for training."
                answer_len = (labels != -100).sum()
                loss_fct = FusedLinearDiffusionCrossEntropyLoss(reduction="sum")
                loss = loss_fct(
                    x=hidden_states,
                    target=labels[logits_to_keep_half].contiguous(),
                    weight=self.lm_head.weight,
                    bias=self.lm_head.bias,
                    p_mask=p_mask_out,
                )
                loss = loss / answer_len
                logits = None

        # ====================== eval / inference ======================
        else:
            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
            hidden_states = hidden_states[:, slice_indices, :].contiguous()

            fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
            if fuse_linear_and_cross_entropy:
                logits = None
            else:
                logits = self.lm_head(hidden_states)

            loss = None
            if labels is not None:
                loss_fct = nn.CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))

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

        if self.training and compute_rl_loss:
            output.entropy = entropy
            output.kl_loss = kl_loss_value if "kl_loss_value" in locals() else torch.tensor(0.0, device=input_ids.device)

        return output


__all__ = [
    "SDARForCausalLM",
    "SDARModel",
    "SDARPreTrainedModel",
]