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.

# 关闭所有 transformers 日志和警告（必须在所有导入前设��）
import os
import sys
import io

os.environ["TRANSFORMERS_VERBOSITY"] = "error"

# 抑制 transformers 和 Triton 的警告日志
class OutputFilter:
    """输出过滤器，抑制特定日志输出，同时保持 tqdm 等工具正常工作"""
    def __init__(self, original_stream):
        self.original_stream = original_stream

    def write(self, text):
        # 过滤规则：
        # 1. 🚨 开头的警告 (transformers docstring)
        # 2. Config not found 消息
        # 3. Triton autotune 日志 (BLOCKS_ARE_CONTIGUOUS, triton_flex_attention, AUTOTUNE benchmarking)
        filtered_keywords = [
            "🚨",
            "Config not found for sdar",
            "BLOCKS_ARE_CONTIGUOUS",
            "triton_flex_attention",
            "AUTOTUNE benchmarking",
        ]
        should_filter = any(keyword in text for keyword in filtered_keywords)
        if not should_filter:
            self.original_stream.write(text)
        return len(text) if not should_filter else 0

    def flush(self):
        self.original_stream.flush()

    # 代理 tqdm 需要的所有属性
    def isatty(self):
        return self.original_stream.isatty()

    @property
    def encoding(self):
        return self.original_stream.encoding

    @property
    def mode(self):
        return self.original_stream.mode

    @property
    def name(self):
        return self.original_stream.name

    def fileno(self):
        return self.original_stream.fileno()

    def __getattr__(self, name):
        return getattr(self.original_stream, name)

# 同时过滤 stdout 和 stderr (Triton 日志打印到 stdout)
sys.stdout = OutputFilter(sys.stdout)
sys.stderr = OutputFilter(sys.stderr)

import logging
import warnings
warnings.filterwarnings("ignore", message=".*is part of.*")
warnings.filterwarnings("ignore", message=".*docstring.*")
# 关闭 torch._inductor autotune 日志
logging.getLogger("torch._inductor").setLevel(logging.CRITICAL)
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           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

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, can_return_tuple, logging
from .configuration_sdar import SDARConfig

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

# ===== FusedLinearDiffusionCrossEntropyLoss =====
try:
    from .fused_linear_diffusion_cross_entropy import FusedLinearDiffusionCrossEntropyLoss
    fused_diffusion_loss_available = True
except ImportError:
    fused_diffusion_loss_available = False
# ===== End FusedLoss =====

import torch.nn.functional as F
# Flash Attention for inference (optional - only required for rollout/eval)
flash_attn_func = None
flash_attn_varlen_func = None
index_first_axis = None
pad_input = None
unpad_input = None
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 ImportError:
    pass

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

# ===== Flex Attention =====
# ===== Flex Attention =====
# Direct import - will error if not available
from torch.nn.attention.flex_attention import BlockMask, create_block_mask, flex_attention
# ===== End Flex Attention =====


logger = logging.get_logger(__name__)



@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


# ===== DiRL Block Diffusion Functions =====
def calculate_token_nums(position_ids: torch.Tensor):
    """
    Calculate token counts per segment for block attention mask creation.

    Args:
        position_ids: (B, L) position ids

    Returns:
        list[list[torch.Tensor]]: nested list of segment lengths for each sample
    """
    if position_ids.dim() != 2:
        raise ValueError(f"Input must be 2D Tensor, got {position_ids.dim()}D")

    all_lengths = []

    for pids_row in position_ids:
        seq_len = pids_row.shape[0]

        # Find indices where value is 0 (segment boundaries)
        zero_indices = torch.nonzero(pids_row == 0).flatten()

        # Add total sequence length as final split point
        split_points = torch.cat([
            zero_indices,
            torch.tensor([seq_len], device=pids_row.device, dtype=zero_indices.dtype)
        ])

        # Calculate differences between adjacent split points
        lengths = torch.diff(split_points)
        all_lengths.append(lengths)

    return all_lengths
# ===== End DiRL Block Diffusion Functions =====


# ===== Flex Attention Block Mask Functions =====
def block_diff_mask(b, h, q_idx, kv_idx, block_size=None, n=None):
    """Construct block diffusion attention mask for flex_attention.

    Three masks combined:
    - Block Diagonal (M_BD): Self-attention within blocks
    - Offset Block-Causal (M_OBC): Cross-attention for context
    - Block-Causal (M_BC): Causal attention within x0

    Args:
        b, h: Batch and head indices (ignored for mask logic).
        q_idx, kv_idx: Query and Key indices.
        block_size: Defines the block structure.
        n: Boundary point for xt/x0 split.

    Returns:
        A boolean attention mask where True = attend, False = mask.
    """
    x0_flag_q = q_idx >= n
    x0_flag_kv = kv_idx >= n

    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
    )

    block_diagonal = (block_q == block_kv) & (x0_flag_q == x0_flag_kv)
    offset_block_causal = (block_q > block_kv) & (x0_flag_kv == 1) & (x0_flag_q == 0)
    block_causal = (block_q >= block_kv) & (x0_flag_kv == 1) & (x0_flag_q == 1)

    return block_diagonal | offset_block_causal | block_causal


def block_attn_mask(num_tokens, block_size, device):
    """Create block attention mask for flex_attention.

    Args:
        num_tokens: list of lists, each inner list is token counts per sample
        block_size: block size for block attention
        device: torch device

    Returns:
        bool tensor (B, N, N) where True = attend
    """
    masks = []
    for i in range(len(num_tokens)):
        cur_masks = []
        for num in num_tokens[i]:
            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))
    return torch.stack(masks, dim=0)


@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
def fused_flex_attention(query, key, value, attention_mask, **kwargs):
    """Compiled flex_attention wrapper for performance."""
    return flex_attention(query, key, value, block_mask=attention_mask, **kwargs)
# ===== End Flex Attention Block Mask Functions =====


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)

        '''
        attention_mask = attention_mask.bool() if attention_mask is not None else 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
            )

        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 = attn_output.transpose(1, 2).contiguous()
        '''

        #print(query_states.shape, key_states.shape, value_states.shape)

        # --- After RoPE and KV-cache handling, expand KV to all heads ---
        key_states   = repeat_kv(key_states,   self.num_key_value_groups)  # [B, H, K, D]
        value_states = repeat_kv(value_states, self.num_key_value_groups)  # [B, H, K, D]

        bsz, q_len = input_shape

        # ===== Attention with Training/Inference Branch =====
        attn_weights = None

        if self.training:
            # ===== Training Mode: Use flex_attention with BlockMask =====
            if isinstance(attention_mask, BlockMask):
                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:
                raise TypeError(
                    f"Expected BlockMask during training, got {type(attention_mask)}. "
                    "Ensure create_flex_block_mask() is called in training code."
                )
        else:
            # ===== Inference Mode: Compatible with regular bool attention_mask =====
            # Used for rollout/eval where tokenizer's attention_mask is passed
            attention_mask_bool = attention_mask.bool() if attention_mask is not None else None

            if attention_mask_bool is not None and torch.all(attention_mask_bool):
                # Full attention (decoding) - use Flash Attention
                if flash_attn_func is None:
                    raise ImportError(
                        "Flash Attention is required for inference. "
                        "Please install flash-attn: pip install flash-attn --no-build-isolation"
                    )
                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:
                # Partial attention (prefilling) - use SDPA
                attn_output = F.scaled_dot_product_attention(
                    query=query_states,
                    key=key_states,
                    value=value_states,
                    attn_mask=attention_mask_bool,
                    is_causal=False,
                    scale=self.scaling,
                    enable_gqa=True
                )
                attn_output = rearrange(attn_output, 'b h l d -> b l (h d)')
        # ===== End Attention =====

        attn_output = self.o_proj(attn_output)
        return attn_output, 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


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)


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
    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:
        r"""
        Args:
            store_kv (`bool`, *optional*):
                Whether to store key and value states in the cache. If `True`, the key and value states will be 
                stored in `past_key_values`. If `False` and `past_key_values` exists, only retrieves cached states 
                without updating the cache.
        """
        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

        # 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):
    ...


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):
        """
        Prepare block diffusion training data (aligned with DiRL).

        Args:
            inputs_ids: (B, L) input token ids
            position_ids: (B, L) position ids
            prompt_mask: (B, L) bool tensor, marks prompt positions
            masked_indices: (B, L) bool tensor, marks which positions are masked
            p_mask_input: (B, L) bool tensor, for RL training p_mask

        Returns:
            concat_inputs_ids: (B, 2*L) concatenated input ids
            concat_position_ids: (B, 2*L) concatenated position ids
            flex_attention_mask_3d: BlockMask for flex_attention
            logits_to_keep_half: (B, L) bool tensor
            logits_to_keep: (B, 2*L) bool tensor
            p_mask: (M,) float tensor
            p_to_keep: (M,) bool tensor or None
        """
        bsz, seq_len = inputs_ids.shape
        num_tokens = calculate_token_nums(position_ids)

        # Get mask_token_id from config, or use default value (bos_token_id + 26, same as DiRL)
        mask_token_id = getattr(self.config, 'mask_token_id', None)
        if mask_token_id is None:
            mask_token_id = self.config.bos_token_id + 26

        # Get block_size from config, or use default value (4, same as DiRL)
        block_size = getattr(self.config, 'block_size', 4)

        # Process masked_indices
        if masked_indices is not None:
            # RL training mode: use provided masked_indices
            noisy_inputs_ids = torch.where(masked_indices, mask_token_id, inputs_ids)
            logits_to_keep_half = masked_indices
            M = masked_indices.sum().item()
            p_mask = torch.full((M,), 0.5, device=inputs_ids.device, dtype=torch.float)
        else:
            # No masked_indices provided: create response_mask (labels != -100 positions)
            if prompt_mask is not None:
                response_mask = (prompt_mask == False)
            else:
                # If no prompt_mask, assume all are response
                response_mask = torch.ones(bsz, seq_len, dtype=torch.bool, device=inputs_ids.device)

            masked_indices = response_mask
            noisy_inputs_ids = torch.where(masked_indices, mask_token_id, inputs_ids)
            logits_to_keep_half = masked_indices
            M = masked_indices.sum().item()
            p_mask = torch.full((M,), 0.5, device=inputs_ids.device, dtype=torch.float)

        # 计算 p_to_keep - 用于将原始空间的 p_mask 映射到提取空间
        p_to_keep = None
        if p_mask_input is not None:
            p_to_keep = p_mask_input[logits_to_keep_half]

        # Build router_noisy_part (marks which positions are noisy)
        # True = original token, False = noisy token, alternating pattern
        # 学习DiRL方式：基于每个样本的segment数量构建
        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)
        concat_inputs_ids = inputs_ids.repeat(1, 2)
        logits_to_keep = torch.zeros(bsz, 2 * seq_len, dtype=torch.bool, device=inputs_ids.device)
        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 (using existing block_attn_mask)
        attention_mask = block_attn_mask(num_tokens, 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
    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,
        # ===== DiRL compatibility parameters =====
        masked_indices: Optional[torch.Tensor] = None,  # (B, L) bool tensor, marks masked positions
        return_logits: bool = False,                    # If True, return only logits for masked positions
        # ===== RL training parameters (compute_rl_loss mode) =====
        compute_rl_loss: bool = False,
        rl_p_mask: Optional[torch.Tensor] = None,
        rl_adv: Optional[torch.Tensor] = None,
        rl_adv_clip_enabled: bool = False,
        rl_adv_clip_min: float = 0.0,
        rl_logp_old_tok: Optional[torch.Tensor] = None,
        rl_logp_ref_tok: Optional[torch.Tensor] = None,
        rl_logp_ref_mask: Optional[torch.Tensor] = None,  # Mask for valid logp_ref positions
        rl_is_real: Optional[torch.Tensor] = None,
        rl_correctness: Optional[torch.Tensor] = None,  # (B,) bool tensor for NLL loss
        rl_ppo_eps: float = 0.2,
        rl_ppo_eps_high: float = 0.28,     # clip eps for advantage actions (adv > 0)
        rl_kl_beta: float = 0.0,           # KL penalty coefficient
        rl_nll_weight: float = 1.0,         # NLL loss coefficient
        rl_use_kl_estimator_k3: bool = True,
        rl_return_entropy: bool = False,
        rl_loss_reduction_mode: str = "token",  # "token" or "sequence"
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> CausalLMOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
        Example:
        ```python
        >>> from transformers import AutoTokenizer, SDARForCausalLM
        >>> model = SDARForCausalLM.from_pretrained("DiffuOpen/SDAR-1.7B-Chat")
        >>> tokenizer = AutoTokenizer.from_pretrained("DiffuOpen/SDAR-1.7B-Chat")
        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")
        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        # ===== DiRL Block Diffusion Training Mode =====
        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

            (
                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=rl_p_mask
            )

            outputs = self.model(
                input_ids=concat_inputs_ids,
                attention_mask=flex_attention_mask_3d,
                position_ids=concat_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
            # Only keep positions marked by logits_to_keep (masked_indices positions)
            hidden_states = hidden_states[logits_to_keep].contiguous()

            # Initialize entropy (in training mode)
            entropy = torch.tensor(0.0, device=input_ids.device)

            # ===== Branch 1: compute_rl_loss (PPO training) =====
            if compute_rl_loss:
                assert p_to_keep is not None, "p_to_keep must be provided for RL loss computation."
                assert rl_adv is not None, "rl_adv must be provided for RL loss computation."
                assert rl_is_real is not None, "rl_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) - M is the number of extracted positions after block diffusion
                logits = self.lm_head(hidden_states)

                # mask — is_real filtering
                is_real_tensor = (
                    rl_is_real.to(device=device, dtype=torch.bool)
                    if torch.is_tensor(rl_is_real)
                    else torch.tensor(rl_is_real, dtype=torch.bool, device=device)
                )
                p_mask_real = rl_p_mask & is_real_tensor.unsqueeze(1)  # (B, L)

                # Compute sequence lengths and batch indices for sequence-level reduction
                if rl_loss_reduction_mode == "sequence":
                    # Count non-padding tokens per sample (B,)
                    seq_lengths = (labels != -100).sum(dim=1)
                    # Create batch index mapping for extracted tokens
                    batch_indices = torch.arange(input_ids.size(0), device=device).unsqueeze(1).expand_as(rl_p_mask)  # (B, L)

                # 正确的索引映射：将 (B, L) 空间的 p_mask_real 映射到提取空间 (M,)
                # logits_to_keep_half (即 masked_indices) 标记了原始空间中哪些位置被提取
                # 在 prepare_for_bd_training 中，提取后的位置对应 logits_to_keep_half 中为 True 的位置
                p_to_keep_real = logits_to_keep_half[p_mask_real]  # (M,) bool

                # Select logits using p_to_keep (在正确的提取空间)
                logits_p = logits[p_to_keep_real]  # (N, V)
                N = p_to_keep_real.sum().item()

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

                # Compute entropy if requested
                entropy_value = None
                if rl_return_entropy:
                    probs_p = log_probs_p.exp()  # Convert log_probs to probs
                    entropy_p = -(probs_p * log_probs_p).sum(dim=-1)  # (N,) per-token entropy
                    entropy_value = entropy_p.mean().detach().clone()  # Scalar mean entropy
                    del probs_p, entropy_p  # Cleanup

                # labels / logp - 需要同样映射到提取空间
                labels_p = labels[p_mask_real][logits_to_keep_half[p_mask_real]]  # (N,)
                logp_p = log_probs_p.gather(dim=-1, index=labels_p.unsqueeze(-1)).squeeze(-1)

                # Extract batch indices for sequence-level grouping
                if rl_loss_reduction_mode == "sequence":
                    batch_indices_p = batch_indices[p_mask_real][logits_to_keep_half[p_mask_real]]  # (N,)

                # Valid token mask (TrajRL: no specific token filtering for now)
                valid_token_mask = torch.ones_like(labels_p, dtype=torch.bool)

                # advantage (TrajRL: per-token advantage, shape (B, L))
                adv_tensor = rl_adv.to(device) if torch.is_tensor(rl_adv) else torch.tensor(rl_adv, dtype=torch.float, device=device)
                # Handle different adv shapes:
                # - (B, L): per-token advantage (TrajRL GAE), use directly
                # - (B,): per-sample advantage (DiRL-style), expand to (B, L)
                if adv_tensor.dim() == 2:
                    adv_expanded = adv_tensor
                elif adv_tensor.dim() == 1:
                    adv_expanded = adv_tensor.unsqueeze(1).expand_as(rl_p_mask)
                else:
                    raise ValueError(f"Unexpected adv shape: {adv_tensor.shape}, expected (B, L) or (B,)")
                # 映射到提取空间
                adv_p = adv_expanded[p_mask_real][logits_to_keep_half[p_mask_real]]

                # old logp - 需要映射到提取空间
                if rl_logp_old_tok is not None and rl_logp_old_tok.numel() > 0:
                    logp_old_p = rl_logp_old_tok.to(device)[p_mask_real][logits_to_keep_half[p_mask_real]]
                else:
                    logp_old_p = logp_p.detach()

                # NLL Loss for correct samples (TrajRL VAPO)
                nll_loss = torch.tensor(0.0, device=device)
                if rl_correctness is not None and rl_correctness.any():
                    correctness_tensor = rl_correctness.to(device=device, dtype=torch.bool)
                    # Expand (B,) to (B, L) following the same pattern as rl_adv (1D case)
                    correctness_expanded = correctness_tensor.unsqueeze(1).expand_as(rl_p_mask)  # (B, L)
                    # 映射到提取空间
                    correct_p = correctness_expanded[p_mask_real][logits_to_keep_half[p_mask_real]]  # (N,)
                    if correct_p.any():
                        nll_loss = rl_nll_weight * (-logp_p[correct_p].mean())

                # KL Loss (aligned with DiRL)
                # Only compute KL for tokens with valid logp_ref (masked out positions use no KL penalty)
                kl_loss = torch.tensor(0.0, device=device)
                if rl_kl_beta > 0 and rl_logp_ref_tok is not None:
                    # 映射到提取空间
                    logp_ref_p = rl_logp_ref_tok.to(device)[p_mask_real][logits_to_keep_half[p_mask_real]]

                    # Apply logp_ref_mask if provided: only compute KL where logp_ref is valid
                    if rl_logp_ref_mask is not None:
                        logp_ref_mask_p = rl_logp_ref_mask.to(device)[p_mask_real][logits_to_keep_half[p_mask_real]]
                        valid_token_mask = valid_token_mask & logp_ref_mask_p

                    # Filter to valid positions only
                    logp_ref_p = logp_ref_p[valid_token_mask]
                    logp_p_valid = logp_p[valid_token_mask]

                    # Compute KL only on valid positions
                    kl_seq_p = logp_p_valid - logp_ref_p

                    if rl_use_kl_estimator_k3:
                        kl_seq_p = (-kl_seq_p).clamp(-10.0, 10.0).exp() - 1.0 + kl_seq_p

                    kl_loss = rl_kl_beta * kl_seq_p.mean()

                # PPO loss with asymmetric clipping (DiRL-style)
                # Use different eps for advantage (adv > 0) vs disadvantage (adv < 0) actions
                ratio_p = (logp_p - logp_old_p).clamp(-10.0, 10.0).exp()

                # Asymmetric clipping based on advantage sign
                # For adv >= 0: use eps_high (more conservative)
                # For adv < 0: use eps (more aggressive)
                clipped_low = ratio_p.clamp(1 - rl_ppo_eps, 1 + rl_ppo_eps)
                clipped_high = ratio_p.clamp(1 - rl_ppo_eps, 1 + rl_ppo_eps_high)
                clipped = torch.where(adv_p >= 0, clipped_high, clipped_low)

                surrogate_p = torch.minimum(ratio_p * adv_p, clipped * adv_p)

                if rl_loss_reduction_mode == "sequence":
                    # Sequence-level reduction: average within each sequence, then average across sequences
                    B = input_ids.size(0)
                    seq_loss_sum = torch.zeros(B, device=device, dtype=surrogate_p.dtype)
                    seq_loss_sum.scatter_add_(0, batch_indices_p, surrogate_p)
                    # Normalize by sequence length
                    seq_loss_mean = seq_loss_sum / seq_lengths.float()
                    # Average across sequences (only non-zero entries)
                    valid_seqs = (seq_lengths > 0)
                    policy_loss = -seq_loss_mean[valid_seqs].mean()
                    # Store sequence-level statistics for logging
                    seq_loss_std = seq_loss_mean[valid_seqs].std()
                    seq_loss_min = seq_loss_mean[valid_seqs].min()
                    seq_loss_max = seq_loss_mean[valid_seqs].max()
                else:
                    # Token-level reduction (default): average across all tokens
                    policy_loss = -surrogate_p.mean()

                # Validation: ensure loss is finite
                if not torch.isfinite(policy_loss):
                    logger.warning(f"Non-finite policy_loss detected: {policy_loss.item()}, mode={rl_loss_reduction_mode}")
                    policy_loss = torch.tensor(0.0, device=device, requires_grad=True)

                # Total loss
                loss = policy_loss + kl_loss + nll_loss

                # Save loss values for output (before cleanup)
                nll_loss_value = nll_loss.detach().clone()
                kl_loss_value = kl_loss.detach().clone()

                # Save PPO statistics for logging (before cleanup)
                # ratio_mean: mean importance sampling ratio
                ratio_mean_value = ratio_p.mean().detach().clone()
                # clip_frac: fraction of samples that were clipped
                clip_frac_value = (ratio_p != clipped).float().mean().detach().clone()

                # Store entropy value (computed earlier)
                entropy_value_stored = entropy_value if entropy_value is not None else torch.tensor(0.0, device=device)

                # Cleanup
                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, policy_loss, entropy_value

                logits = None

            # ===== Branch 2: return_logits (inference) =====
            elif return_logits:
                logits = self.lm_head(hidden_states)  # (M, V)
                loss = None

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

            # ===== Branch 3: block diffusion loss (pre-training) =====
            else:
                assert labels is not None, "Labels must be provided for training."

                # Use FusedLinearDiffusionCrossEntropyLoss if available
                if fused_diffusion_loss_available and getattr(self.config, 'use_fused_diffusion_loss', False):
                    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=getattr(self.lm_head, 'bias', None),
                        p_mask=p_mask_out,
                    )
                    loss = loss / answer_len
                else:
                    # Fallback: CrossEntropyLoss on full logits
                    logits_full = self.lm_head(hidden_states)
                    loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
                    loss = loss_fct(logits_full.view(-1, self.config.vocab_size), labels.view(-1))

                logits = None

            # ===== Block Diffusion Training Mode: Construct output directly =====
            # Construct output and return directly (don't execute Standard Mode code)
            output = CausalLMOutputWithPast(
                loss=loss,
                logits=logits,
                past_key_values=outputs.past_key_values,
                hidden_states=outputs.hidden_states,
                attentions=outputs.attentions,
            )

            # Add RL loss attributes for training (DiRL-style)
            if self.training and compute_rl_loss:
                output.nll_loss = nll_loss_value if "nll_loss_value" in locals() else torch.tensor(0.0, device=input_ids.device)
                output.kl_loss = kl_loss_value if "kl_loss_value" in locals() else torch.tensor(0.0, device=input_ids.device)
                # Add PPO statistics for logging
                output.ratio_mean = ratio_mean_value if "ratio_mean_value" in locals() else torch.tensor(0.0, device=input_ids.device)
                output.clip_frac = clip_frac_value if "clip_frac_value" in locals() else torch.tensor(0.0, device=input_ids.device)
                output.entropy = entropy_value_stored if "entropy_value_stored" in locals() else torch.tensor(0.0, device=input_ids.device)
                output.seq_loss_std = seq_loss_std if "seq_loss_std" in locals() else torch.tensor(0.0, device=input_ids.device)
                output.seq_loss_min = seq_loss_min if "seq_loss_min" in locals() else torch.tensor(0.0, device=input_ids.device)
                output.seq_loss_max = seq_loss_max if "seq_loss_max" in locals() else torch.tensor(0.0, device=input_ids.device)

            return output

        else:
            # ===== Standard Mode (no block diffusion) =====
            # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
            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
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        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 = getattr(self.config, 'fuse_cross_entropy', False) and self.training
        if fuse_linear_and_cross_entropy:
            # When using fused_linear_ce_loss, we do not compute the whole logits on HBM
            logits = None
        else:
            logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # ===== FusedLinearDiffusionCrossEntropyLoss (if available) =====
            use_fused_loss = (
                fused_diffusion_loss_available and
                getattr(self.config, 'use_fused_diffusion_loss', False) and
                self.training
            )

            if use_fused_loss:
                # FusedLoss directly computes from hidden_states, no need for logits
                p_mask = kwargs.get('p_mask', None)
                if p_mask is None:
                    # Default: all ones (no noise scaling)
                    # Match the shape of labels for correct reshaping in FusedLoss
                    if labels.dim() >= 2:
                        p_mask = torch.ones_like(labels, dtype=hidden_states.dtype, device=hidden_states.device)
                    else:
                        # labels is already 1D
                        p_mask = torch.ones(labels.shape, dtype=hidden_states.dtype, device=hidden_states.device)

                num_chunks = getattr(self.config, 'fused_loss_num_chunks', 8)
                loss_fct = FusedLinearDiffusionCrossEntropyLoss(
                    ignore_index=-100,
                    label_smoothing=0.0,
                    logit_scale=1.0,
                    num_chunks=num_chunks,
                    reduction="mean"
                )
                loss = loss_fct(
                    x=hidden_states,
                    target=labels,
                    weight=self.lm_head.weight,
                    bias=getattr(self.lm_head, 'bias', None),
                    p_mask=p_mask,
                )
            else:
                # Standard CrossEntropyLoss (fallback)
                loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
                loss = loss_fct(
                    logits.view(-1, self.config.vocab_size), labels.view(-1))
            # ===== End FusedLoss =====


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

        # Add RL loss attributes for training (DiRL-style)
        if self.training and compute_rl_loss:
            output.nll_loss = nll_loss_value if "nll_loss_value" in locals() else torch.tensor(0.0, device=input_ids.device)
            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",
]