modeling.py

from dataclasses import dataclass
from typing import Any, Callable, Optional, Union

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

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, ModelOutput
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 auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
from transformers.utils.deprecation import deprecate_kwarg
from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm
from .configuration import Fast_dDriveConfig, Fast_dDriveTextConfig, Fast_dDriveVisionConfig

from torch.nn.attention.flex_attention import flex_attention, create_block_mask, or_masks

from functools import partial
import random
import math

# Context-parallel (CP) and Section-MoE-LoRA are research-only extensions that
# are not part of the canonical paper release. We provide no-op stubs so the
# original call sites behave as if running on a single-process, no-LoRA setup.
def is_cp_enabled():
    return False


def get_cp_rank():
    return 0


def get_cp_size():
    return 1


def get_cp_group():
    return None


def cp_allgather_kv(*args, **kwargs):
    raise RuntimeError("cp_allgather_kv called but context-parallel is disabled in the release build")


def rewrite_mask_mod_for_cp(mask_mod, *args, **kwargs):
    return mask_mod


def shard_doubled_sequence(x, *args, **kwargs):
    return x


def shard_position_ids(x, *args, **kwargs):
    return x


def _set_section_ids(*args, **kwargs):
    pass


def _build_section_ids_tensor(*args, **kwargs):
    return None


from .section_utils import (
    compute_section_block_idx_deep_static as _compute_section_block_idx_deep_static,
    build_deep_scaffold_sequences as _build_deep_scaffold_sequences,
    NULL_TOKEN_ID as _NULL_TOKEN_ID,
)

logger = logging.get_logger(__name__)


# flex_attention MUST be torch.compiled to use the efficient Triton sparse
# kernel; without it the call falls back to eager dense O(Q×KV) attention.
# TORCHDYNAMO_DISABLE=1 (often set in training scripts to avoid full-model
# compilation overhead) would silently turn @torch.compile into a no-op,
# so we temporarily re-enable dynamo for these two definitions only.
_dynamo_was_disabled = getattr(torch._dynamo.config, "disable", False)
if _dynamo_was_disabled:
    torch._dynamo.config.disable = False

@torch.compile()
def fused_flex_attention(q, k, v, mask=None):
    return flex_attention(q, k, v, block_mask=mask, enable_gqa=True)

# Compiled flex_attention for quadratic speculative decoding hot path.
# enable_gqa=True avoids repeat_interleave.
_compiled_flex_attention_quadratic = torch.compile(flex_attention)

if _dynamo_was_disabled:
    torch._dynamo.config.disable = True

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_causal_mask(b, h, q_idx, kv_idx, block_size=None, n=None):

    # 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 = (q_idx >= kv_idx) & (x0_flag_kv == 1) & (x0_flag_q == 1)

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


def hybrid_block_causal_mask_multiturn(b, h, q_idx, kv_idx, response_block_idx=None, turn_idx=None, n=None):
    """
    Multi-turn hybrid mask: Prompt uses causal, Response uses block causal.
    
    Args:
        response_block_idx: [seq_len] tensor, -1 for prompt, >=0 for response block index
        turn_idx: [seq_len] tensor, turn index for each position (0, 1, 2, ...)
        n: sequence length (half of total)
    
    Rules:
    - Each token can see all previous turns
    - Within current turn: prompt uses causal, response uses block causal
    - x_t response sees x_0: only tokens from current turn and before
    - x_0: standard causal mask
    
    Example for [prompt1, response1, prompt2, response2]:
    - prompt1 (turn 0): causal within turn 0 prompt
    - response1 (turn 0): sees prompt1 + block causal within response1
    - prompt2 (turn 1): sees all of turn 0 + causal within turn 1 prompt
    - response2 (turn 1): sees all of turn 0 + prompt2 + block causal within response2
    """
    x0_flag_q = (q_idx >= n)
    x0_flag_kv = (kv_idx >= n)
    
    pos_q = torch.where(x0_flag_q, q_idx - n, q_idx)
    pos_kv = torch.where(x0_flag_kv, kv_idx - n, kv_idx)
    
    block_q = response_block_idx[pos_q]
    block_kv = response_block_idx[pos_kv]
    turn_q = turn_idx[pos_q]
    turn_kv = turn_idx[pos_kv]
    
    is_prompt_q = (block_q < 0)
    is_prompt_kv = (block_kv < 0)
    
    # x_t region rules:
    # 1. Can see all previous turns: turn_q > turn_kv
    # 2. Within same turn, prompt: causal (turn same + is prompt + pos satisfies causal)
    # 3. Within same turn, response: sees all prompt in same turn + block causal for response
    # xt_same_turn_prompt_causal = ~x0_flag_q & ~x0_flag_kv & (turn_q == turn_kv) & is_prompt_q & (pos_q >= pos_kv)
    # xt_same_turn_response = ~x0_flag_q & ~x0_flag_kv & (turn_q == turn_kv) & ~is_prompt_q & (
    #     ~is_prompt_kv
    # )
    block_diagonal = ~x0_flag_q & ~x0_flag_kv & (turn_q == turn_kv) 
    
    # **2. Offset Block-Causal Mask (M_OBC) **
    offset_block_causal = (
        (turn_q > turn_kv) 
        & (x0_flag_kv == 1)
        & (x0_flag_q == 0)
    )
    # x_0 region: standard causal
    x0_causal = x0_flag_q & x0_flag_kv & (pos_q >= pos_kv)
    
    return (block_diagonal | 
            offset_block_causal | 
            x0_causal)


def eval_block_diff_mask(q_idx, kv_idx, block_size=None):
    # Compute block indices
    block_q = q_idx // block_size
    block_kv = kv_idx // block_size

    return torch.ones_like(block_q >= block_kv)

def eval_causal_mask(q_idx, kv_idx):
    return q_idx >= kv_idx


def eval_hybrid_block_causal_mask(q_idx, kv_idx, response_block_idx):
    """
    Inference-time hybrid block causal mask matching training's
    hybrid_block_causal_mask_multiturn pattern.

    For prompt tokens (block_idx == -1): standard causal mask.
    For response tokens: block-causal — can see all prompt tokens,
    bidirectional within same block, causal across blocks
    (block i can see blocks 0..i but not i+1..N).

    Args:
        q_idx:  [Q, 1] query position indices
        kv_idx: [1, K] key/value position indices
        response_block_idx: [seqlen] tensor, -1 for prompt, >=0 for block index

    Returns:
        [Q, K] boolean mask where True = can attend
    """
    block_q = response_block_idx[q_idx]    # [Q, 1]
    block_kv = response_block_idx[kv_idx]  # [1, K]

    is_prompt_q = (block_q < 0)
    is_prompt_kv = (block_kv < 0)

    # Prompt → prompt: standard causal
    prompt_causal = is_prompt_q & is_prompt_kv & (q_idx >= kv_idx)
    # Response → prompt: can see all prompt tokens
    response_sees_prompt = ~is_prompt_q & is_prompt_kv
    # Response → response: block causal (same block = bidirectional, earlier block = OK)
    response_block_causal = ~is_prompt_q & ~is_prompt_kv & (block_q >= block_kv)

    return prompt_causal | response_sees_prompt | response_block_causal


def _crop_dynamic_cache(past_key_values: DynamicCache, max_length: int):
    """Crop DynamicCache to max_length (used after draft_only phase in quadratic speculative decoding)."""
    new_past = []
    for layer_num in range(len(past_key_values)):
        layer_kv = ()
        for kv_idx in range(len(past_key_values[layer_num])):
            layer_kv += (past_key_values[layer_num][kv_idx][:, :, :max_length, :],)
        new_past.append(layer_kv)
    return DynamicCache(new_past)


def _extract_draft_kv_cache(past_key_values: DynamicCache, clean_len: int, block_length: int):
    """After quadratic decoding, extract only draft tokens (first of each block) from cache."""
    new_past = []
    for layer_num in range(len(past_key_values)):
        layer_kv = ()
        for kv_idx in range(len(past_key_values[layer_num])):
            tensor = past_key_values[layer_num][kv_idx]
            clean_part = tensor[:, :, :clean_len, :]
            draft_part = tensor[:, :, clean_len:: (block_length + 1), :]
            layer_kv += (torch.cat([clean_part, draft_part], dim=2),)
        new_past.append(layer_kv)
    return DynamicCache(new_past)


class Fast_dDriveMLP(nn.Module):
    def __init__(self, config, bias: bool = False):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=bias)
        self.act_fn = ACT2FN[config.hidden_act]

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


class Fast_dDriveVisionPatchEmbed(nn.Module):
    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 2,
        in_channels: int = 3,
        embed_dim: int = 1152,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.in_channels = in_channels
        self.embed_dim = embed_dim

        kernel_size = [temporal_patch_size, patch_size, patch_size]
        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


class Fast_dDriveVisionRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs


class Fast_dDrivePatchMerger(nn.Module):
    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        self.ln_q = Qwen2RMSNorm(context_dim, eps=1e-6)
        self.mlp = nn.Sequential(
            nn.Linear(self.hidden_size, self.hidden_size),
            nn.GELU(),
            nn.Linear(self.hidden_size, dim),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
        return x


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_vision(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
    orig_q_dtype = q.dtype
    orig_k_dtype = k.dtype
    q, k = q.float(), k.float()
    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    q_embed = q_embed.to(orig_q_dtype)
    k_embed = k_embed.to(orig_k_dtype)
    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 Fast_dDriveVisionAttention(nn.Module):
    def __init__(self, config: Fast_dDriveVisionConfig) -> None:
        super().__init__()
        self.dim = config.hidden_size
        self.num_heads = config.num_heads
        self.head_dim = self.dim // self.num_heads
        self.num_key_value_groups = 1  # needed for eager attention
        self.qkv = nn.Linear(self.dim, self.dim * 3, bias=True)
        self.proj = nn.Linear(self.dim, self.dim)
        self.scaling = self.head_dim**-0.5
        self.config = config
        self.attention_dropout = 0.0
        self.is_causal = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        query_states, key_states, value_states = (
            self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        )
        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)

        query_states = query_states.transpose(0, 1).unsqueeze(0)
        key_states = key_states.transpose(0, 1).unsqueeze(0)
        value_states = value_states.transpose(0, 1).unsqueeze(0)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        if self.config._attn_implementation == "flash_attention_2":
            # Flash Attention 2: Use cu_seqlens for variable length attention
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
            attn_output, _ = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask=None,
                scaling=self.scaling,
                dropout=0.0 if not self.training else self.attention_dropout,
                cu_seq_lens_q=cu_seqlens,
                cu_seq_lens_k=cu_seqlens,
                max_length_q=max_seqlen,
                max_length_k=max_seqlen,
                is_causal=False,
                **kwargs,
            )
        else:
            # Other implementations: Process each chunk separately
            lengths = cu_seqlens[1:] - cu_seqlens[:-1]
            splits = [
                torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
            ]

            attn_outputs = [
                attention_interface(
                    self,
                    q,
                    k,
                    v,
                    attention_mask=None,
                    scaling=self.scaling,
                    dropout=0.0 if not self.training else self.attention_dropout,
                    is_causal=False,
                    **kwargs,
                )[0]
                for q, k, v in zip(*splits)
            ]
            attn_output = torch.cat(attn_outputs, dim=1)

        attn_output = attn_output.reshape(seq_length, -1).contiguous()
        attn_output = self.proj(attn_output)
        return attn_output


class Fast_dDriveVisionBlock(GradientCheckpointingLayer):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = Qwen2RMSNorm(config.hidden_size, eps=1e-6)
        self.norm2 = Qwen2RMSNorm(config.hidden_size, eps=1e-6)
        self.attn = Fast_dDriveVisionAttention(config=config)
        self.mlp = Fast_dDriveMLP(config, bias=True)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> torch.Tensor:
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states),
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


@auto_docstring
class Fast_dDrivePreTrainedModel(PreTrainedModel):
    config: Fast_dDriveConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Fast_dDriveDecoderLayer", "Fast_dDriveVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True

    def gradient_checkpointing_enable(
        self,
        gradient_checkpointing_kwargs: Optional[dict[str, Any]] = None,
    ) -> None:
        """
        Ensure non-reentrant checkpointing when the trainers call into Transformers'
        gradient checkpointing helper. Flash attention kernels used by MDM do not
        support reentrant checkpointing, so we request the safer path by default.
        """
        if gradient_checkpointing_kwargs is None:
            gradient_checkpointing_kwargs = {}
        else:
            gradient_checkpointing_kwargs = dict(gradient_checkpointing_kwargs)

        gradient_checkpointing_kwargs.setdefault("use_reentrant", False)
        super().gradient_checkpointing_enable(gradient_checkpointing_kwargs=gradient_checkpointing_kwargs)


class Fast_dDriveVisionTransformerPretrainedModel(Fast_dDrivePreTrainedModel):
    config: Fast_dDriveVisionConfig
    _no_split_modules = ["Fast_dDriveVisionBlock"]

    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = config.patch_size
        self.fullatt_block_indexes = config.fullatt_block_indexes
        self.window_size = config.window_size
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size

        self.patch_embed = Fast_dDriveVisionPatchEmbed(
            patch_size=config.patch_size,
            temporal_patch_size=config.temporal_patch_size,
            in_channels=config.in_channels,
            embed_dim=config.hidden_size,
        )

        head_dim = config.hidden_size // config.num_heads
        self.rotary_pos_emb = Fast_dDriveVisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList([Fast_dDriveVisionBlock(config) for _ in range(config.depth)])
        self.merger = Fast_dDrivePatchMerger(
            dim=config.out_hidden_size,
            context_dim=config.hidden_size,
            spatial_merge_size=config.spatial_merge_size,
        )
        self.gradient_checkpointing = False

    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def get_window_index(self, grid_thw):
        window_index: list = []
        cu_window_seqlens: list = [0]
        window_index_id = 0
        vit_merger_window_size = self.window_size // self.spatial_merge_size // self.patch_size

        for grid_t, grid_h, grid_w in grid_thw:
            llm_grid_h, llm_grid_w = (
                grid_h // self.spatial_merge_size,
                grid_w // self.spatial_merge_size,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )
            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
        window_index = torch.cat(window_index, dim=0)

        return window_index, cu_window_seqlens

    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
                The final hidden states of the model.
            grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
                The temporal, height and width of feature shape of each image in LLM.

        Returns:
            `torch.Tensor`: hidden_states.
        """
        hidden_states = self.patch_embed(hidden_states)
        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        window_index, cu_window_seqlens = self.get_window_index(grid_thw)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=hidden_states.device,
            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        hidden_states = hidden_states[window_index, :, :]
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0,
            # Select dtype based on the following factors:
            #  - FA2 requires that cu_seqlens_q must have dtype int32
            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
            # See https://github.com/huggingface/transformers/pull/34852 for more information
            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

        for layer_num, blk in enumerate(self.blocks):
            if layer_num in self.fullatt_block_indexes:
                cu_seqlens_now = cu_seqlens
            else:
                cu_seqlens_now = cu_window_seqlens

            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens_now,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.merger(hidden_states)
        reverse_indices = torch.argsort(window_index)
        hidden_states = hidden_states[reverse_indices, :]

        return hidden_states


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Llava outputs, with hidden states and attentions.
    """
)
class Fast_dDriveModelOutputWithPast(ModelOutput):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
        The rope index difference between sequence length and multimodal rope.
    """

    last_hidden_state: Optional[torch.FloatTensor] = None
    past_key_values: Optional[list[torch.FloatTensor]] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class Fast_dDriveRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Fast_dDriveTextConfig, 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()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        # In contrast to other models, Fast_dDrive has different position ids for the grids
        # So we expand the inv_freq to shape (3, ...)
        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)

        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(2, 3)
            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 Qwen2MLP(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):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).

    Explanation:
        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
        vision embedding part, we apply rotary position embedding on temporal, height and width dimension separately.
        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
        difference with modern LLMs.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        mrope_section(`List(int)`):
            Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    mrope_section = mrope_section * 2
    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )
    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )

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


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

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

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.is_causal = True
        self.attention_dropout = config.attention_dropout
        self.rope_scaling = config.rope_scaling
        self.scaling = self.head_dim**-0.5

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None

        self.rotary_emb = Fast_dDriveRotaryEmbedding(config=config)
        self._quadratic_block_mask: dict = {}

    def _get_sbd_inference_quadratic_decoding_block_mask(self, block_length: int):
        """Build block mask for quadratic speculative decoding (one forward for full block)."""
        if block_length not in self._quadratic_block_mask:
            draft_len = block_length * (block_length + 1)

            def quadratic(b, h, q_idx, kv_idx):
                first_clean = torch.logical_and(
                    kv_idx % (block_length + 1) == 0,
                    kv_idx < draft_len,
                )
                first_clean = torch.logical_and(first_clean, q_idx >= kv_idx)
                block_q = q_idx // (block_length + 1)
                block_kv = kv_idx // (block_length + 1)
                same_block = torch.logical_and(block_q == block_kv, q_idx < draft_len)
                same_block_except_first = torch.logical_and(
                    same_block,
                    q_idx % (block_length + 1) != 0,
                )
                draft_part = torch.logical_or(first_clean, same_block_except_first)
                clean_part = kv_idx >= draft_len
                return torch.logical_or(draft_part, clean_part)

            block_mask = create_block_mask(
                quadratic,
                B=None,
                H=None,
                Q_LEN=draft_len,
                KV_LEN=draft_len + self.config.max_position_embeddings,
                device="cuda",
            )
            self._quadratic_block_mask[block_length] = block_mask
        return self._quadratic_block_mask[block_length]

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        update_kv_cache: bool = False,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

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

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

        cos, sin = position_embeddings
        if self.training:
            #split q into two parts
            q_1 = query_states[:,:,:query_states.shape[2]//2]
            q_2 = query_states[:,:,query_states.shape[2]//2:]
            #split k into two parts
            k_1 = key_states[:,:,:key_states.shape[2]//2]
            k_2 = key_states[:,:,key_states.shape[2]//2:]
            q_1, k_1 = apply_multimodal_rotary_pos_emb(q_1, k_1, cos, sin, self.rope_scaling["mrope_section"])
            q_2, k_2 = apply_multimodal_rotary_pos_emb(q_2, k_2, cos, sin, self.rope_scaling["mrope_section"])
            query_states = torch.cat((q_1, q_2), dim=-2)
            key_states = torch.cat((k_1, k_2), dim=-2)
        else:
            query_states, key_states = apply_multimodal_rotary_pos_emb(
                query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
            )

        if past_key_values is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
            if update_kv_cache:
                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
            # elif len(past_key_values) > self.layer_idx:
            elif len(past_key_values) > self.layer_idx and past_key_values[self.layer_idx][0] is not None:
                key_states = torch.cat((past_key_values[self.layer_idx][0], key_states), dim=-2)
                value_states = torch.cat((past_key_values[self.layer_idx][1], value_states), dim=-2)

        self_spec_mode = getattr(self.config, "self_spec_inference_mode", None)
        block_length = getattr(self.config, "block_length", None) or getattr(self.config, "bd_size", None)
        if not self.training and self_spec_mode is not None and block_length is not None:
            if self_spec_mode == "quadratic" and past_key_values is not None:
                # HOT PATH: main loop of quadratic speculative decoding.
                # Use compiled flex_attention + enable_gqa=True (no repeat_interleave)
                # for Triton sparse kernel instead of eager O(Q×KV) dense fallback.
                seq_len = key_states.shape[2]
                draft_len = block_length * (block_length + 1)
                clean_keys = key_states[:, :, :-draft_len]
                draft_keys = key_states[:, :, -draft_len:]
                clean_values = value_states[:, :, :-draft_len]
                draft_values = value_states[:, :, -draft_len:]
                key_states = torch.cat([draft_keys, clean_keys], dim=2)
                value_states = torch.cat([draft_values, clean_values], dim=2)
                block_mask = self._get_sbd_inference_quadratic_decoding_block_mask(block_length)
                block_mask.seq_lengths = (draft_len, seq_len)
                attn_output = _compiled_flex_attention_quadratic(
                    query_states, key_states, value_states,
                    block_mask=block_mask, enable_gqa=True,
                )
            else:
                # COLD PATH: draft_only ("default") or non-cached quadratic.
                # Called once per sample — eager flex_attention is acceptable.
                key_states = key_states.repeat_interleave(self.num_key_value_groups, dim=1)
                value_states = value_states.repeat_interleave(self.num_key_value_groups, dim=1)
                if self_spec_mode == "quadratic":
                    # Non-cached quadratic (initial forward without past_key_values)
                    seq_len = query_states.shape[2]
                    draft_len = block_length * (block_length + 1)
                    clean_len = seq_len - draft_len

                    def _causal_mask(b, h, q_idx, kv_idx):
                        return torch.logical_and(q_idx >= kv_idx, q_idx < clean_len)

                    def _draft2clean_mask(b, h, q_idx, kv_idx):
                        full_clean = torch.logical_and(q_idx >= clean_len, kv_idx < clean_len)
                        first_clean = torch.logical_and(
                            q_idx >= clean_len,
                            (kv_idx - clean_len) % (block_length + 1) == 0,
                        )
                        first_clean = torch.logical_and(first_clean, q_idx >= kv_idx)
                        return torch.logical_or(full_clean, first_clean)

                    def _draft_mask(b, h, q_idx, kv_idx):
                        block_q = (q_idx - clean_len) // (block_length + 1)
                        block_kv = (kv_idx - clean_len) // (block_length + 1)
                        quadrant = torch.logical_and(q_idx >= clean_len, kv_idx >= clean_len)
                        same_block = torch.logical_and(block_q == block_kv, quadrant)
                        same_block_except_first = torch.logical_and(
                            same_block,
                            (q_idx - clean_len) % (block_length + 1) != 0,
                        )
                        return torch.logical_and(same_block, same_block_except_first)

                    mask = or_masks(_causal_mask, _draft2clean_mask)
                    mask = or_masks(mask, _draft_mask)
                    block_mask = create_block_mask(
                        mask, B=None, H=None, Q_LEN=seq_len, KV_LEN=seq_len,
                    )
                else:
                    # self_spec_mode == "default": clean causal + draft sees all
                    seq_len = query_states.shape[2]
                    prefix_len = seq_len - block_length

                    def _clean_q_mask(b, h, q_idx, kv_idx):
                        return torch.logical_and(q_idx >= kv_idx, q_idx < prefix_len)

                    def _noisy_q_mask(b, h, q_idx, kv_idx):
                        return q_idx >= prefix_len

                    block_mask = create_block_mask(
                        or_masks(_clean_q_mask, _noisy_q_mask),
                        B=None,
                        H=None,
                        Q_LEN=seq_len,
                        KV_LEN=seq_len,
                    )
                attn_output = flex_attention(query_states, key_states, value_states, block_mask=block_mask)
            attn_output = attn_output.transpose(1, 2).reshape(bsz, q_len, -1).contiguous()
            attn_output = self.o_proj(attn_output)
            return attn_output, None

        if self.training:
            # CP: all-gather KV so each rank sees full sequence
            if is_cp_enabled():
                key_states, value_states = cp_allgather_kv(
                    key_states.contiguous(), value_states.contiguous()
                )

            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

            # RoPE produces fp32 Q,K (cos/sin are fp32) while V stays bf16.
            # flex_attention requires Q,K,V to share a dtype.
            if query_states.dtype != value_states.dtype:
                query_states = query_states.to(value_states.dtype)
                key_states = key_states.to(value_states.dtype)

            attn_output = fused_flex_attention(query_states, key_states, value_states, mask=attention_mask)
            attn_output = attn_output.transpose(1, 2).contiguous()
            attn_weights = None
        else:
            attention_interface: Callable = eager_attention_forward
            if self.config._attn_implementation != "eager":
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

            attn_output, attn_weights = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                sliding_window=self.sliding_window,
                position_ids=position_ids,  # pass positions for FA2
                **kwargs,
            )

        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Fast_dDriveDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Fast_dDriveTextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

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

        self.mlp = Qwen2MLP(config)
        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        update_kv_cache: bool = False,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_values (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            update_kv_cache=update_kv_cache,
            **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 Fast_dDriveTextModel(Fast_dDrivePreTrainedModel):
    config: Fast_dDriveTextConfig

    def __init__(self, config: Fast_dDriveTextConfig):
        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(
            [Fast_dDriveDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Fast_dDriveRotaryEmbedding(config=config)
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

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

    
    @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,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        update_kv_cache: bool = False,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # torch.jit.trace() doesn't support cache objects in the output
        if use_cache and past_key_values is None and not torch.jit.is_tracing():
            past_key_values = DynamicCache(config=self.config)

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

        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
            )

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

        # NOTE: we need to pass text position ids for packing. Qwen2-VL uses 3D positions
        # where each dim indicates visual spatial positions for temporal/height/width grids.
        # There are two scenarios when FA2-like packed masking might be activated.
        # 1. User specifically passed packed `position_ids` and no attention mask.
        #    In this case we expect the useer to create correct position ids for all 3 grids
        #    and prepend text-only position ids to it. The final tensor will be [4, bs, seq-len]
        # 2. User runs forward with no attention mask and no position ids. In this case, position ids
        #    are prepared by the model (`get_rope_index`) as `[4, bs, seq-len]` tensor. Text-only positions are
        #    prepended by us when creating positions so that the mask is constructed correctly. NOTE: failing to pass
        #    text-only positions will cause incorrect mask construction, do not change `prepare_input_for_generation`
        if position_ids.ndim == 3 and position_ids.shape[0] == 4:
            text_position_ids = position_ids[0]
            position_ids = position_ids[1:]
        else:
            # If inputs are not packed (usual 3D positions), do not prepare mask from position_ids
            text_position_ids = None

        # It may already have been prepared by e.g. `generate`
        # if not isinstance(causal_mask_mapping := attention_mask, dict):
        #     # Prepare mask arguments
        #     mask_kwargs = {
        #         "config": self.config,
        #         "input_embeds": inputs_embeds,
        #         "attention_mask": attention_mask,
        #         "cache_position": cache_position,
        #         "past_key_values": past_key_values,
        #         "position_ids": text_position_ids,
        #     }
        #     # Create the masks
        #     causal_mask_mapping = {
        #         "full_attention": create_causal_mask(**mask_kwargs),
        #     }
        #     # The sliding window alternating layers are not always activated depending on the config
        #     if self.has_sliding_layers:
        #         causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)

        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:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=attention_mask.to(device=hidden_states.device),
                position_ids=text_position_ids,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                update_kv_cache=update_kv_cache,
                **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,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns] if v is not None
            )
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


@auto_docstring
class Fast_dDriveModel(Fast_dDrivePreTrainedModel):
    base_model_prefix = ""
    _checkpoint_conversion_mapping = {"^model": "language_model"}
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False
    config: Fast_dDriveConfig
    _no_split_modules = ["Fast_dDriveDecoderLayer", "Fast_dDriveVisionBlock"]

    def __init__(self, config):
        super().__init__(config)
        self.visual = Fast_dDriveVisionTransformerPretrainedModel._from_config(config.vision_config)
        self.language_model = Fast_dDriveTextModel._from_config(config.text_config)
        self.rope_deltas = None  # cache rope_deltas here
        self.use_block_causal_mask = config.use_block_causal_mask

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

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

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

    def get_decoder(self):
        return self.language_model

    def get_rope_index(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Calculate the 3D rope index based on image and video's temporal, height and width in LLM.

        Explanation:
            Each embedding sequence contains vision embedding and text embedding or just contains text embedding.

            For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs.
            Examples:
                input_ids: [T T T T T], here T is for text.
                temporal position_ids: [0, 1, 2, 3, 4]
                height position_ids: [0, 1, 2, 3, 4]
                width position_ids: [0, 1, 2, 3, 4]

            For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
            and 1D rotary position embedding for text part.
            Examples:
                Temporal (Time): 3 patches, representing different segments of the video in time.
                Height: 2 patches, dividing each frame vertically.
                Width: 2 patches, dividing each frame horizontally.
                We also have some important parameters:
                fps (Frames Per Second): The video's frame rate, set to 1. This means one frame is processed each second.
                tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens" are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens per second. So each second of the video will be represented with 25 separate time points. It essentially defines the temporal granularity.
                temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames.
                interval: The step size for the temporal position IDs, calculated as tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50. This means that each temporal patch will be have a difference of 50 in the temporal position IDs.
                input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
                vision temporal position_ids: [0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100]
                vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
                vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
                text temporal position_ids: [101, 102, 103, 104, 105]
                text height position_ids: [101, 102, 103, 104, 105]
                text width position_ids: [101, 102, 103, 104, 105]
                Here we calculate the text start position_ids as the max vision position_ids plus 1.

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
                it.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
            second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
                The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

        Returns:
            position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
            mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
        """
        spatial_merge_size = self.config.vision_config.spatial_merge_size
        image_token_id = self.config.image_token_id
        video_token_id = self.config.video_token_id
        vision_start_token_id = self.config.vision_start_token_id
        mrope_position_deltas = []
        if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
            total_input_ids = input_ids
            if attention_mask is not None:
                attention_mask = attention_mask == 1
            position_ids = torch.ones(
                3,
                input_ids.shape[0],
                input_ids.shape[1],
                dtype=input_ids.dtype,
                device=input_ids.device,
            )
            image_index, video_index = 0, 0
            for i, input_ids in enumerate(total_input_ids):
                if attention_mask is not None:
                    input_ids = input_ids[attention_mask[i]]
                image_nums, video_nums = 0, 0
                vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
                vision_tokens = input_ids[vision_start_indices + 1]
                image_nums = (vision_tokens == image_token_id).sum()
                video_nums = (vision_tokens == video_token_id).sum()
                input_tokens = input_ids.tolist()
                llm_pos_ids_list: list = []
                st = 0
                remain_images, remain_videos = image_nums, video_nums
                if image_nums + video_nums == 0:
                    image_index += 1
                    video_index += 1
                    continue
                for _ in range(image_nums + video_nums):
                    if image_token_id in input_tokens and remain_images > 0:
                        ed_image = input_tokens.index(image_token_id, st)
                    else:
                        ed_image = len(input_tokens) + 1
                    if video_token_id in input_tokens and remain_videos > 0:
                        ed_video = input_tokens.index(video_token_id, st)
                    else:
                        ed_video = len(input_tokens) + 1
                    if ed_image < ed_video:
                        t, h, w = (
                            image_grid_thw[image_index][0],
                            image_grid_thw[image_index][1],
                            image_grid_thw[image_index][2],
                        )
                        second_per_grid_t = 0
                        image_index += 1
                        remain_images -= 1
                        ed = ed_image

                    else:
                        t, h, w = (
                            video_grid_thw[video_index][0],
                            video_grid_thw[video_index][1],
                            video_grid_thw[video_index][2],
                        )
                        if second_per_grid_ts is not None:
                            second_per_grid_t = second_per_grid_ts[video_index]
                        else:
                            second_per_grid_t = 1.0
                        video_index += 1
                        remain_videos -= 1
                        ed = ed_video
                    llm_grid_t, llm_grid_h, llm_grid_w = (
                        t.item(),
                        h.item() // spatial_merge_size,
                        w.item() // spatial_merge_size,
                    )
                    text_len = ed - st

                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                    llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

                    range_tensor = torch.arange(llm_grid_t).view(-1, 1)
                    expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)

                    ## normalize type, send to device.
                    second_per_grid_t = torch.as_tensor(
                        second_per_grid_t, dtype=range_tensor.dtype, device=range_tensor.device
                    )

                    time_tensor = expanded_range * second_per_grid_t * self.config.vision_config.tokens_per_second

                    time_tensor_long = time_tensor.long()
                    t_index = time_tensor_long.flatten()

                    h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
                    w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
                    llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
                    st = ed + llm_grid_t * llm_grid_h * llm_grid_w

                if st < len(input_tokens):
                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                    text_len = len(input_tokens) - st
                    llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

                llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
                if attention_mask is not None:
                    position_ids[..., i, attention_mask[i]] = llm_positions.to(position_ids.device)
                else:
                    position_ids[..., i, :] = llm_positions.to(position_ids.device)
                mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
            mrope_position_deltas = torch.tensor(mrope_position_deltas).unsqueeze(1).to(device=input_ids.device)
            return position_ids, mrope_position_deltas
        else:
            # if attention_mask is not None:
            #     position_ids = attention_mask.long().cumsum(-1) - 1
            #     position_ids.masked_fill_(attention_mask == 0, 1)
            #     position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
            #     max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
            #     mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
            # else:
            if self.training:
                position_ids = (
                    torch.arange(input_ids.shape[1] // 2, device=input_ids.device)
                    .view(1, 1, -1)
                    .expand(3, input_ids.shape[0], -1)
                )
            else:
                if attention_mask is not None:
                    position_ids = (attention_mask.long().cumsum(-1) - 1)[-1]
                    position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
                    max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
                    mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
                else:
                    position_ids = (
                        torch.arange(input_ids.shape[1], device=input_ids.device)
                        .view(1, 1, -1)
                        .expand(3, input_ids.shape[0], -1)
                    )
            mrope_position_deltas = torch.zeros(
                [input_ids.shape[0], 1],
                device=input_ids.device,
                dtype=input_ids.dtype,
            )

            return position_ids, mrope_position_deltas

    def get_video_features(
        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
    ):
        """
        Encodes videos into continuous embeddings that can be forwarded to the language model.

        Args:
            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
                The tensors corresponding to the input videos.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
        """
        pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
        split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        video_embeds = torch.split(video_embeds, split_sizes)
        return video_embeds

    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
        """
        Encodes images into continuous embeddings that can be forwarded to the language model.

        Args:
            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
                The tensors corresponding to the input images.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
        """
        pixel_values = pixel_values.type(self.visual.dtype)
        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        image_embeds = torch.split(image_embeds, split_sizes)
        return image_embeds

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
        image_features: Optional[torch.FloatTensor] = None,
        video_features: Optional[torch.FloatTensor] = None,
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
            special_video_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_video_mask = special_video_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id
            special_video_mask = input_ids == self.config.video_token_id

        n_image_tokens = special_image_mask.sum()
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
            raise ValueError(
                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
            )

        n_video_tokens = special_video_mask.sum()
        special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
            raise ValueError(
                f"Videos features and video tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
            )

        return special_image_mask, special_video_mask


    def eval_mask(self, seqlen, block_size, cache_seq_len, update_kv_cache=False, use_block_causal_mask=False):
        q_indices = torch.arange(seqlen, device=self.device) + cache_seq_len
        k_indices = torch.arange(seqlen + cache_seq_len, device=self.device)
        if use_block_causal_mask and update_kv_cache:
            mask = eval_causal_mask(q_indices[:, None], k_indices[None, :])
        else:
            mask = eval_block_diff_mask(
                q_idx=q_indices[:, None], 
                kv_idx=k_indices[None, :], 
                block_size=block_size
            )
        return mask

    def eval_hybrid_mask(self, seqlen, response_block_idx):
        """Build inference-time hybrid block causal mask from response_block_idx.

        Matches the training-time hybrid_block_causal_mask_multiturn pattern:
        prompt = causal, response = block-causal with section-aware block indices.

        Args:
            seqlen: total sequence length
            response_block_idx: [seqlen] tensor, -1 for prompt, >=0 for block index

        Returns:
            [seqlen, seqlen] boolean mask
        """
        q_indices = torch.arange(seqlen, device=self.device)
        k_indices = torch.arange(seqlen, device=self.device)
        mask = eval_hybrid_block_causal_mask(
            q_idx=q_indices[:, None],
            kv_idx=k_indices[None, :],
            response_block_idx=response_block_idx,
        )
        return mask

    @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,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        update_kv_cache: bool = False,
        bd_size: Optional[int] = None,
        **kwargs,
    ) -> Union[tuple, Fast_dDriveModelOutputWithPast]:
        r"""
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
            The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

        if pixel_values is not None:
            image_embeds = self.get_image_features(pixel_values, image_grid_thw)
            image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            image_mask, _ = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

        if pixel_values_videos is not None:
            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
            video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            _, video_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

        if position_ids is None:
            # Calculate RoPE index once per generation in the pre-fill stage only.
            # When compiling, we can't check tensor values thus we check only input length
            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
            # models currently cannot do asssisted decoding
            prefill_compiled_stage = is_torchdynamo_compiling() and (
                (input_ids is not None and input_ids.shape[1] != 1)
                or (inputs_embeds is not None and inputs_embeds.shape[1] != 1)
            )
            prefill_noncompiled_stage = not is_torchdynamo_compiling() and (
                (cache_position is not None and cache_position[0] == 0)
                or (past_key_values is None or past_key_values.get_seq_length() == 0)
            )
            if (prefill_compiled_stage or prefill_noncompiled_stage) or self.rope_deltas is None:
                position_ids, rope_deltas = self.get_rope_index(
                    input_ids,
                    image_grid_thw,
                    video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    attention_mask=attention_mask,
                )
                self.rope_deltas = rope_deltas
            else:
                batch_size, seq_length, _ = inputs_embeds.shape

                if self.training and pixel_values is None and pixel_values_videos is None: # only train on text
                    seq_length = seq_length // 2

                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
                # if cache_position is not None:
                #     delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
                if past_key_values is not None:
                    delta = (past_key_values.get_seq_length() + self.rope_deltas).to(inputs_embeds.device)
                else:
                    delta = torch.zeros((batch_size, seq_length), device=inputs_embeds.device)
                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=1)
                position_ids = position_ids + delta.to(position_ids.device)

        position_ids = position_ids.to(inputs_embeds.device)
        if not self.training:
            if attention_mask is None:
                attention_mask = self.eval_mask(inputs_embeds.shape[1], self.bd_size if bd_size is None else bd_size, 0 if past_key_values is None else past_key_values.get_seq_length(), update_kv_cache=update_kv_cache, use_block_causal_mask=self.use_block_causal_mask).to(inputs_embeds.device)
            # else: use pre-computed attention_mask (e.g. from mdm_sample_deep_scaffold)

        outputs = self.language_model(
            input_ids=None,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            update_kv_cache=update_kv_cache,
            **kwargs,
        )

        output = Fast_dDriveModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=self.rope_deltas,
        )
        return output if return_dict else output.to_tuple()


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Fast_dDrive causal language model (or autoregressive) outputs.
    """
)
class Fast_dDriveCausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
        The rope index difference between sequence length and multimodal rope.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: Optional[torch.FloatTensor] = None
    past_key_values: Optional[list[torch.FloatTensor]] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class Fast_dDriveForConditionalGeneration(Fast_dDrivePreTrainedModel, GenerationMixin):
    config_class = Fast_dDriveConfig
    _checkpoint_conversion_mapping = {
        "^visual": "model.visual",
        r"^model(?!\.(language_model|visual))": "model.language_model",
    }
    _tied_weights_keys = ["lm_head.weight"]
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = True

    @classmethod
    def from_pretrained(cls, *args, **kwargs):
        # HF's ``modeling_utils.PreTrainedModel.from_pretrained`` only applies
        # ``_checkpoint_conversion_mapping`` when the class name contains one of
        # the substrings in the built-in ``VLMS`` allow-list (e.g. ``qwen2_5_vl``).
        # Our class name (``Fast_dDriveForConditionalGeneration``) does not match,
        # so we pass the mapping explicitly here.  Without this, weights for the
        # ``model.language_model.*`` and ``model.visual.*`` subtrees would
        # silently stay on the meta device.
        if "key_mapping" not in kwargs:
            kwargs["key_mapping"] = cls._checkpoint_conversion_mapping
        return super().from_pretrained(*args, **kwargs)

    def __init__(self, config):
        super().__init__(config)
        self.model = Fast_dDriveModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.bd_size = config.bd_size
        self.model.bd_size = self.bd_size
        self.complementary_mask = getattr(config, 'complementary_mask', False)
        self.always_mask_im_end = getattr(config, 'always_mask_im_end', False)
        self.flexible_bd_size = getattr(config, 'flexible_bd_size', False)
        self.use_block_causal_mask = getattr(config, 'use_block_causal_mask', False)
        self.anneal_block_size = getattr(config, 'anneal_block_size', False)
        self.enable_efficient_vision_embed = getattr(config, 'enable_efficient_vision_embed', False)
        self.minimum_noise_level = getattr(config, 'minimum_noise_level', 0.0)
        self.entropy_loss = getattr(config, 'entropy_loss', False)
        self.entropy_loss_weight = getattr(config, 'entropy_loss_weight', 1.0)
        self.block_causal_no_dynamic = getattr(config, 'block_causal_no_dynamic', False)
        self.use_json_scaffold = getattr(config, 'use_json_scaffold', False)
        self.section_token_budgets = getattr(config, 'section_token_budgets', None)
        self.deep_json_scaffold = getattr(config, 'deep_json_scaffold', False)
        self._section_tokenizer = None  # lazily set from outside
        self._deep_scaffold_sequences = None  # lazily built on first forward
        self.null_token_id = _NULL_TOKEN_ID
        self.im_end_token_id = 151645  # <|im_end|> token id
        # SASD: Section-Aware Semantics Diffusion
        self.section_loss_weights = getattr(config, 'section_loss_weights', None)
        self.section_noise_schedule = getattr(config, 'section_noise_schedule', None)
        # Section-MoE-LoRA (set to True externally by finetune_qwenvl.py)
        self._use_section_moe_lora = False
        # self.max_context_length = 4096
        
        # Vision-to-text aligner (if vision output dim != text hidden dim)
        vision_out_dim = config.vision_config.out_hidden_size
        text_hidden = config.text_config.hidden_size
        if vision_out_dim != text_hidden:
            self.vision_to_text_proj = nn.Linear(vision_out_dim, text_hidden, bias=False)
            for p in self.vision_to_text_proj.parameters():
                p.requires_grad = False
            logger.info(f"Vision-to-text aligner: {vision_out_dim} -> {text_hidden}")
        else:
            self.vision_to_text_proj = None
        
        self.post_init()

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

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

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

    def get_decoder(self):
        return self.model.get_decoder()

    def get_video_features(
        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
    ):
        return self.model.get_video_features(pixel_values_videos, video_grid_thw)

    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
        return self.model.get_image_features(pixel_values, image_grid_thw)

    # Make modules available through conditional class for BC
    @property
    def language_model(self):
        return self.model.language_model

    @property
    def visual(self):
        return self.model.visual

    def gen_mask(self, seqlen, block_size, B, H):
        # ================== 修改开始 ==================
        # flex_attention 要求闭包捕获的变量必须是 Tensor
        # 将 int 转换为 Tensor，并放在对应的设备上
        block_size_t = torch.tensor(block_size, device=self.device, dtype=torch.int32)
        n_t = torch.tensor(seqlen, device=self.device, dtype=torch.int32)

        mask_fn = partial(block_diff_mask, block_size=block_size_t, n=n_t)

        if is_cp_enabled():
            local_half = seqlen // get_cp_size()
            mask_fn = rewrite_mask_mod_for_cp(
                mask_fn, get_cp_rank(), local_half, seqlen, self.device
            )
            mask = create_block_mask(
                mask_fn, B=B, H=H,
                Q_LEN=2 * local_half, KV_LEN=2 * seqlen
            )
        else:
            mask = create_block_mask(
                mask_fn, B=B, H=H,
                Q_LEN=seqlen * 2, KV_LEN=seqlen * 2
            )
        # ================== 修改结束 ==================
        return mask

    def gen_block_causal_mask(self, seqlen, block_size, B, H):
        block_size_t = torch.tensor(block_size, device=self.device, dtype=torch.int32)
        n_t = torch.tensor(seqlen, device=self.device, dtype=torch.int32)
        mask_fn = partial(block_causal_mask, block_size=block_size_t, n=n_t)

        if is_cp_enabled():
            local_half = seqlen // get_cp_size()
            mask_fn = rewrite_mask_mod_for_cp(
                mask_fn, get_cp_rank(), local_half, seqlen, self.device
            )
            mask = create_block_mask(
                mask_fn, B=B, H=H,
                Q_LEN=2 * local_half, KV_LEN=2 * seqlen
            )
        else:
            mask = create_block_mask(
                mask_fn, B=B, H=H,
                Q_LEN=seqlen * 2, KV_LEN=seqlen * 2
            )
        return mask

    def compute_response_block_idx(self, labels, block_size):
        """
        Compute block index and turn index for each position. 
        Each response segment has independent blocks.
        
        Example: prompt1(3) + response1(14) + prompt2(2) + response2(2)
        - turn_idx: [0,0,0, 0,0,...,0, 1,1, 1,1] (prompt+response = same turn)
        - response1: 14 tokens → 2 blocks (0, 1) with sizes (8, 6)
        - response2: 2 tokens → 1 block (2) with size (2)
        - Total: 3 blocks
        
        Returns:
            response_block_idx: [seq_len] where prompt=-1, response=block_idx
            turn_idx: [seq_len] turn index for each position
            n_blocks: total number of blocks
        """
        labels_single = labels[0]  # [seq_len]
        seq_len = labels_single.shape[0]
        response_mask = (labels_single != -100)
        
        response_block_idx = torch.full((seq_len,), -1, device=labels.device, dtype=torch.int64)
        turn_idx = torch.zeros((seq_len,), device=labels.device, dtype=torch.int64)
        
        current_block = 0
        in_response = False
        response_pos_in_segment = 0  # position within current response segment
        
        for i in range(seq_len):
            if response_mask[i]:
                if not in_response:
                    # Start of new response segment
                    in_response = True
                    response_pos_in_segment = 0
                
                # Block index within this segment + global offset
                block_in_segment = response_pos_in_segment // block_size
                response_block_idx[i] = current_block + block_in_segment
                response_pos_in_segment += 1
            else:
                if in_response:
                    # End of response segment, update current_block and start new turn
                    n_blocks_in_segment = (response_pos_in_segment + block_size - 1) // block_size
                    current_block += n_blocks_in_segment
                    in_response = False
        
        for i in range(1, seq_len):
            if response_block_idx[i] != response_block_idx[i-1]:
                turn_idx[i] = turn_idx[i-1] + 1
            else:
                turn_idx[i] = turn_idx[i-1]
                
        # Handle case where sequence ends with response
        if in_response:
            n_blocks_in_segment = (response_pos_in_segment + block_size - 1) // block_size
            current_block += n_blocks_in_segment
        
        n_blocks = current_block
        return response_block_idx, turn_idx, n_blocks

    def gen_hybrid_block_causal_mask(self, seqlen, response_block_idx, turn_idx, B, H):
        """Generate hybrid mask: prompt causal, response block causal."""
        n_t = torch.tensor(seqlen, device=self.device, dtype=torch.int32)
        mask_fn = partial(
            hybrid_block_causal_mask_multiturn,
            response_block_idx=response_block_idx, turn_idx=turn_idx, n=n_t,
        )

        if is_cp_enabled():
            local_half = seqlen // get_cp_size()
            mask_fn = rewrite_mask_mod_for_cp(
                mask_fn, get_cp_rank(), local_half, seqlen, self.device
            )
            mask = create_block_mask(
                mask_fn, B=B, H=H,
                Q_LEN=2 * local_half, KV_LEN=2 * seqlen
            )
        else:
            mask = create_block_mask(
                mask_fn, B=B, H=H,
                Q_LEN=seqlen * 2, KV_LEN=seqlen * 2
            )
        return mask

    def _sample_section_aware_noise(self, n_blocks, device, block_to_section=None):
        """Sample noise levels t for each block using section-specific Beta distributions.

        Falls back to uniform for blocks not assigned to any section.
        ``block_to_section`` is provided by ``compute_section_block_idx_deep_static``.
        """
        if block_to_section is None:
            block_to_section = {}
        t = torch.zeros(n_blocks, device=device)
        for block_idx in range(n_blocks):
            section_name = block_to_section.get(block_idx)
            if section_name is not None and self.section_noise_schedule is not None and section_name in self.section_noise_schedule:
                alpha_str, beta_str = self.section_noise_schedule[section_name].split(",")
                alpha, beta = float(alpha_str), float(beta_str)
                t[block_idx] = torch.distributions.Beta(alpha, beta).sample().to(device)
            else:
                t[block_idx] = torch.rand(1, device=device).item()
        return t

    def _build_section_weight_tensor(self, labels, response_block_idx, n_blocks, block_to_section=None):
        """Build a per-token weight tensor based on section loss weights.

        Returns:
            weight_tensor: [B, seq_len] with section weights for each token
        """
        if block_to_section is None:
            block_to_section = {}
        weight_tensor = torch.ones_like(labels, dtype=torch.float32)

        for i in range(labels.shape[1]):
            block_idx = response_block_idx[i].item() if response_block_idx.dim() == 1 else response_block_idx[i].item()
            if block_idx >= 0:
                section_name = block_to_section.get(block_idx)
                if section_name is not None and self.section_loss_weights is not None and section_name in self.section_loss_weights:
                    weight_tensor[:, i] = self.section_loss_weights[section_name]

        return weight_tensor

    def compute_section_weighted_loss(self, logits, labels, weight_tensor, vocab_size, num_items_in_batch=None):
        """Compute cross-entropy loss with per-token section weights.

        Args:
            logits: [B, seq_len, vocab_size]
            labels: [B, seq_len] with -100 for ignored tokens
            weight_tensor: [B, seq_len] with section weights
            vocab_size: vocabulary size
            num_items_in_batch: for gradient accumulation normalization
        """
        # Shift for causal LM
        shift_logits = logits[..., :-1, :].contiguous()
        shift_labels = labels[..., 1:].contiguous()
        shift_weights = weight_tensor[..., 1:].contiguous()

        # Flatten
        shift_logits = shift_logits.view(-1, vocab_size)
        shift_labels = shift_labels.view(-1)
        shift_weights = shift_weights.view(-1)

        # Per-token CE (no reduction)
        loss_fct = nn.CrossEntropyLoss(reduction='none', ignore_index=-100)
        per_token_loss = loss_fct(shift_logits, shift_labels)

        # Apply section weights
        weighted_loss = per_token_loss * shift_weights

        # Reduce
        if num_items_in_batch is not None:
            return weighted_loss.sum() / num_items_in_batch
        else:
            non_ignore = shift_labels != -100
            if non_ignore.sum() == 0:
                return torch.tensor(0.0, device=logits.device)
            return weighted_loss[non_ignore].sum() / non_ignore.sum()

    def compute_entropy_loss(self, logits, labels, num_items_in_batch=None):
        """Compute entropy loss with optional global normalization.
        
        Args:
            logits: Model logits
            labels: Ground truth labels (-100 for ignored tokens)
            num_items_in_batch: Global number of non-ignored tokens for normalization.
                               If provided, uses sum/num_items_in_batch for global norm.
                               If None, uses mean() for micro-batch norm.
        """
        non_ignore_mask = labels != -100
        logits = logits[non_ignore_mask]
        labels = labels[non_ignore_mask]
        correct_mask = logits.argmax(dim=-1) == labels

        compute_logits = logits[correct_mask]

        if correct_mask.sum() == 0:
            return torch.tensor(0.0, device=logits.device)

        p = F.softmax(compute_logits, dim=-1)
        log_p = F.log_softmax(compute_logits, dim=-1)
        entropy = -torch.sum(p * log_p, dim=-1)

        if num_items_in_batch is not None:
            # Global normalization: use same denominator as cross entropy loss
            return entropy.sum() / num_items_in_batch
        else:
            return entropy.mean()

    @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,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        mask_id: Optional[int] = 151665,
        update_kv_cache: bool = False,
        eval_bd_size: Optional[int] = None,
        **kwargs,
    ) -> Union[tuple, Fast_dDriveCausalLMOutputWithPast]:
        # input_ids = torch.tensor([[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]]).to(input_ids.device, dtype=input_ids.dtype)
        # labels = torch.tensor([[-100,-100,3,4,5,6,-100,-100,-100,-100,11,12,13,14,15]]).to(labels.device, dtype=labels.dtype)
        # pixel_values = None
        # pixel_values_videos = None
        # self.bd_size = 2
        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:
            if self.anneal_block_size:
                # Get update_ratio from kwargs (passed by trainer)
                update_ratio = kwargs.get('update_ratio', 1.0)
                # Compute possible bd_sizes: [2, 4, 8, ..., target_bd_size]
                max_power = int(math.log2(self.bd_size))
                possible_bd_sizes = [2**i for i in range(2, max_power + 1)]  # Start from 4
                # Select bd_size based on update_ratio
                idx = min(int(update_ratio * len(possible_bd_sizes)), len(possible_bd_sizes) - 1)
                bd_size = possible_bd_sizes[idx]
            elif self.flexible_bd_size:
                max_power = int(math.log2(self.bd_size))
                possible_bd_sizes = [2**i for i in range(max_power + 1)]
                bd_size = random.choice(possible_bd_sizes)
            else:
                bd_size = self.bd_size
            import time as _time, os as _os
            _debug_fwd = _os.environ.get("DVLM_DEBUG_FORWARD", "0") == "1"

            if pixel_values is None and pixel_values_videos is None: # only train on text

                batch_size, seq_len = input_ids.shape
                original_labels = labels.clone()
                original_input_ids = input_ids.clone()

                # Compute response block index: -1 for prompt, >=0 for response
                # Each response segment has independent blocks
                response_mask = (labels != -100)  # [B, seq_len]
                eps = self.minimum_noise_level

                # Section-aware variable block sizes (text-only path)
                _scaffold_mask_text = None
                _block_to_section = None  # populated by deep scaffold for SASD
                if self.deep_json_scaffold and self._section_tokenizer is not None:
                    # Deep scaffold: freeze sub-keys within sections
                    if self._deep_scaffold_sequences is None:
                        self._deep_scaffold_sequences = _build_deep_scaffold_sequences(self._section_tokenizer)
                    response_block_idx, turn_idx, n_blocks, _scaffold_mask_text, _block_to_section = _compute_section_block_idx_deep_static(
                        labels, original_input_ids,
                        self._deep_scaffold_sequences,
                        fallback_block_size=bd_size,
                    )
                # Section-aware mask sampling (deep scaffold v2 only)
                if self.deep_json_scaffold and self._section_tokenizer is not None:
                    # SASD: section-adaptive noise schedule
                    if self.section_noise_schedule is not None:
                        t = self._sample_section_aware_noise(n_blocks, input_ids.device, block_to_section=_block_to_section)
                    else:
                        t = torch.rand((n_blocks,), device=input_ids.device)
                    p_mask_per_block = (1 - eps) * t + eps

                    mask_indices = torch.zeros_like(labels, dtype=torch.bool)
                    for i in range(seq_len):
                        block_i = response_block_idx[i].item()
                        if block_i >= 0:
                            mask_indices[:, i] = torch.rand((batch_size,), device=input_ids.device) < p_mask_per_block[block_i]

                    # Freeze scaffold tokens (never mask them)
                    if self.use_json_scaffold and _scaffold_mask_text is not None:
                        mask_indices = mask_indices & ~_scaffold_mask_text.unsqueeze(0).expand_as(mask_indices)

                elif self.use_block_causal_mask and not self.block_causal_no_dynamic:
                    response_block_idx, turn_idx, n_blocks = self.compute_response_block_idx(labels, bd_size)

                    # Sample t for each block: [n_blocks]

                    # random sample t for each block from [self.minimum_noise_level, 1]
                    t = torch.rand((n_blocks,), device=input_ids.device)
                    p_mask_per_block = (1 - eps) * t + eps

                    # Create mask_indices: [B, seq_len]
                    mask_indices = torch.zeros_like(labels, dtype=torch.bool)
                    for i in range(seq_len):
                        block_i = response_block_idx[i].item()
                        if block_i >= 0:  # response token
                            mask_indices[:, i] = torch.rand((batch_size,), device=input_ids.device) < p_mask_per_block[block_i]
                else:
                    input_ids = input_ids.reshape(input_ids.shape[0] * input_ids.shape[1] // bd_size, bd_size)
                    b, l = input_ids.shape
                    t = torch.rand((b,), device=input_ids.device)
                    p_mask = (1 - eps) * t + eps
                    p_mask = p_mask[:, None].repeat(1, l)

                    mask_indices = torch.rand((b, l), device=input_ids.device) < p_mask
                    mask_indices = mask_indices.reshape(labels.shape) & response_mask
                    input_ids = input_ids.reshape(labels.shape)

                # Always mask <|im_end|> in response
                if self.always_mask_im_end:
                    im_end_mask = (input_ids == self.im_end_token_id) & response_mask
                    mask_indices = mask_indices | im_end_mask

                # Apply mask only to response
                noisy_input_ids = input_ids.clone()
                noisy_input_ids[mask_indices] = mask_id

                # Update labels: only predict masked response tokens
                labels = labels.clone()
                labels[~mask_indices] = -100

                # Concatenate [noisy | clean]
                input_ids = torch.cat([noisy_input_ids, original_input_ids], dim=1)

                # Complementary version
                if self.complementary_mask:
                    complementary_mask_indices = response_mask & ~mask_indices
                    if self.always_mask_im_end:
                        im_end_mask = (original_input_ids == self.im_end_token_id) & response_mask
                        complementary_mask_indices = complementary_mask_indices | im_end_mask

                    # Also freeze scaffold in complementary mask
                    if self.use_json_scaffold and _scaffold_mask_text is not None:
                        complementary_mask_indices = complementary_mask_indices & ~_scaffold_mask_text.unsqueeze(0).expand_as(complementary_mask_indices)

                    complementary_noisy_input_ids = original_input_ids.clone()
                    complementary_noisy_input_ids[complementary_mask_indices] = mask_id

                    complementary_labels = original_labels.clone()
                    complementary_labels[~complementary_mask_indices] = -100
                    complementary_input_ids = torch.cat([complementary_noisy_input_ids, original_input_ids], dim=1)

                    input_ids = torch.cat([input_ids, complementary_input_ids], dim=0)
                    labels = torch.cat([labels, complementary_labels], dim=0)

                if self.deep_json_scaffold and self._section_tokenizer is not None:
                    # Deep scaffold v2: always uses hybrid block causal mask
                    attention_mask = self.gen_hybrid_block_causal_mask(seq_len, response_block_idx, turn_idx, input_ids.shape[0], self.config.num_attention_heads)
                elif self.use_block_causal_mask:
                    if self.block_causal_no_dynamic:
                        attention_mask = self.gen_block_causal_mask(seq_len, bd_size, input_ids.shape[0], self.config.num_attention_heads)
                    else:
                        attention_mask = self.gen_hybrid_block_causal_mask(seq_len, response_block_idx, turn_idx, input_ids.shape[0], self.config.num_attention_heads)
                else:
                    attention_mask = self.gen_mask(seq_len, bd_size, input_ids.shape[0], self.config.num_attention_heads)
        
            else:  # 多模态 block diffusion
                # Phase A: Embed + masked scatter vision
                import time as _time, os as _os
                _debug_fwd = _os.environ.get("DVLM_DEBUG_FORWARD", "0") == "1"
                if _debug_fwd:
                    torch.cuda.synchronize()
                    _t_phaseA = _time.perf_counter()

                if inputs_embeds is None:
                    inputs_embeds = self.model.get_input_embeddings()(input_ids)
                
                if pixel_values is not None:
                    image_embeds = self.model.get_image_features(pixel_values, image_grid_thw)
                    image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
                    if self.vision_to_text_proj is not None:
                        image_embeds = self.vision_to_text_proj(image_embeds)
                    image_mask, _ = self.model.get_placeholder_mask(
                        input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
                    )
                    inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
                
                if pixel_values_videos is not None:
                    video_embeds = self.model.get_video_features(pixel_values_videos, video_grid_thw)
                    video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
                    if self.vision_to_text_proj is not None:
                        video_embeds = self.vision_to_text_proj(video_embeds)
                    _, video_mask = self.model.get_placeholder_mask(
                        input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
                    )
                    inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
                
                # Phase B: 生成 3D position_ids（在扩倍前，基于原长）
                if position_ids is None:
                    position_ids, rope_deltas = self.model.get_rope_index(
                        input_ids=input_ids,
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
                        second_per_grid_ts=second_per_grid_ts,
                        attention_mask=attention_mask,
                    )
                
                # Phase C: Block diffusion (保护 vision token 位置)
                if _debug_fwd:
                    torch.cuda.synchronize()
                    _t_phaseAB = _time.perf_counter()
                batch_size = input_ids.shape[0]
                L = input_ids.shape[1]
                seq_len = L

                # if L > self.max_context_length:
                #     L = self.max_context_length
                #     input_ids = input_ids[:, :self.max_context_length]
                #     labels = labels[:, :self.max_context_length]
                #     position_ids = position_ids[:, :self.max_context_length]
                #     attention_mask = attention_mask[:, :self.max_context_length]
                #     inputs_embeds = inputs_embeds[:, :self.max_context_length]

                hidden_size = inputs_embeds.shape[-1]
                
                original_labels = labels.clone()
                original_input_ids = input_ids.clone()
                original_embeds = inputs_embeds.clone()
                original_position_ids = position_ids.clone()  # 保存原长 position [3, B, L]
                
                # 识别 vision tokens（不加噪声）
                image_token_id = self.config.image_token_id
                video_token_id = self.config.video_token_id
                vision_start_token_id = self.config.vision_start_token_id
                vision_token_mask = (input_ids == image_token_id) | (input_ids == video_token_id) | (input_ids == vision_start_token_id)
                vision_mask_3d = vision_token_mask.unsqueeze(-1).expand(-1, -1, hidden_size)
                
                # Block diffusion with multi-turn support
                # Each response segment has independent blocks
                response_mask = (labels != -100)  # [B, seq_len]
                eps = self.minimum_noise_level

                # Section-aware variable block sizes (multimodal path)
                _scaffold_mask_mm = None
                _block_to_section = None  # populated by deep scaffold for SASD
                if self.deep_json_scaffold and self._section_tokenizer is not None:
                    # Deep scaffold: freeze sub-keys within sections
                    if self._deep_scaffold_sequences is None:
                        self._deep_scaffold_sequences = _build_deep_scaffold_sequences(self._section_tokenizer)
                    response_block_idx, turn_idx, n_blocks, _scaffold_mask_mm, _block_to_section = _compute_section_block_idx_deep_static(
                        labels, original_input_ids,
                        self._deep_scaffold_sequences,
                        fallback_block_size=bd_size,
                    )
                # Section-aware mask sampling (deep scaffold v2 only)
                if self.deep_json_scaffold and self._section_tokenizer is not None:
                    # SASD: section-adaptive noise schedule
                    if self.section_noise_schedule is not None:
                        t = self._sample_section_aware_noise(n_blocks, input_ids.device, block_to_section=_block_to_section)
                    else:
                        t = torch.rand((n_blocks,), device=input_ids.device)
                    p_mask_per_block = (1 - eps) * t + eps

                    mask_indices = torch.zeros_like(labels, dtype=torch.bool)
                    for i in range(seq_len):
                        block_i = response_block_idx[i].item()
                        if block_i >= 0:
                            mask_indices[:, i] = torch.rand((batch_size,), device=input_ids.device) < p_mask_per_block[block_i]

                    # Freeze scaffold tokens (never mask them)
                    if self.use_json_scaffold and _scaffold_mask_mm is not None:
                        mask_indices = mask_indices & ~_scaffold_mask_mm.unsqueeze(0).expand_as(mask_indices)

                elif self.use_block_causal_mask and not self.block_causal_no_dynamic:
                    response_block_idx, turn_idx, n_blocks = self.compute_response_block_idx(labels, bd_size)

                    # Sample t for each block: [n_blocks]

                    # random sample t for each block from [self.minimum_noise_level, 1]
                    t = torch.rand((n_blocks,), device=input_ids.device)
                    p_mask_per_block = (1 - eps) * t + eps

                    # Create mask_indices: [B, seq_len]
                    mask_indices = torch.zeros_like(labels, dtype=torch.bool)
                    for i in range(seq_len):
                        block_i = response_block_idx[i].item()
                        if block_i >= 0:  # response token
                            mask_indices[:, i] = torch.rand((batch_size,), device=input_ids.device) < p_mask_per_block[block_i]
                else:
                    response_block_idx, turn_idx, n_blocks = self.compute_response_block_idx(labels, bd_size)
                    input_ids = input_ids.reshape(input_ids.shape[0] * input_ids.shape[1] // bd_size, bd_size)
                    b, l = input_ids.shape
                    t = torch.rand((b,), device=input_ids.device)
                    p_mask = (1 - eps) * t + eps
                    p_mask = p_mask[:, None].repeat(1, l)

                    mask_indices = torch.rand((b, l), device=input_ids.device) < p_mask
                    mask_indices = mask_indices.reshape(labels.shape) & response_mask
                    input_ids = input_ids.reshape(labels.shape)

                if self.always_mask_im_end:
                    im_end_mask = (input_ids == self.im_end_token_id) & response_mask
                    mask_indices = mask_indices | im_end_mask

                noisy_input_ids = input_ids.clone()
                noisy_input_ids[mask_indices] = mask_id

                # Noisy embeds（保护 vision）
                if self.enable_efficient_vision_embed:
                    noisy_embeds = original_embeds.clone()
                    text_mask_3d = mask_indices.unsqueeze(-1).expand(-1, -1, hidden_size)
                    mask_embeds = self.model.language_model.embed_tokens(
                        torch.full_like(input_ids, mask_id)
                    )
                    noisy_embeds = torch.where(text_mask_3d, mask_embeds, noisy_embeds)
                else:
                    noisy_embeds_raw = self.model.language_model.embed_tokens(noisy_input_ids)
                    noisy_embeds = torch.where(vision_mask_3d, original_embeds, noisy_embeds_raw)

                # 更新 labels
                labels_noisy = labels.clone()
                labels_noisy[~mask_indices] = -100

                # 拼接 [noisy | clean]
                input_ids_pair1 = torch.cat([noisy_input_ids, original_input_ids], dim=1)
                embeds_pair1 = torch.cat([noisy_embeds, original_embeds], dim=1)
                labels_pair1 = labels_noisy
                position_ids_pair1 = original_position_ids  # [3, B, L]

                input_ids = input_ids_pair1
                inputs_embeds = embeds_pair1
                labels = labels_pair1
                position_ids = position_ids_pair1

                # Complementary
                if self.complementary_mask:
                    complementary_mask_indices = response_mask & ~mask_indices
                    if self.always_mask_im_end:
                        im_end_mask = (original_input_ids == self.im_end_token_id) & response_mask
                        complementary_mask_indices = complementary_mask_indices | im_end_mask

                    # Also freeze scaffold in complementary mask
                    if self.use_json_scaffold and _scaffold_mask_mm is not None:
                        complementary_mask_indices = complementary_mask_indices & ~_scaffold_mask_mm.unsqueeze(0).expand_as(complementary_mask_indices)

                    complementary_noisy_input_ids = original_input_ids.clone()
                    complementary_noisy_input_ids[complementary_mask_indices] = mask_id

                    if self.enable_efficient_vision_embed:
                        complementary_noisy_embeds = original_embeds.clone()
                        text_mask_3d = complementary_mask_indices.unsqueeze(-1).expand(-1, -1, hidden_size)
                        mask_embeds = self.model.language_model.embed_tokens(
                            torch.full_like(original_input_ids, mask_id)
                        )
                        complementary_noisy_embeds = torch.where(text_mask_3d, mask_embeds, complementary_noisy_embeds)
                    else:
                        complementary_noisy_embeds_raw = self.model.language_model.embed_tokens(complementary_noisy_input_ids)
                        complementary_noisy_embeds = torch.where(vision_mask_3d, original_embeds, complementary_noisy_embeds_raw)

                    complementary_labels = original_labels.clone()
                    complementary_labels[~complementary_mask_indices] = -100
                    
                    input_ids_pair2 = torch.cat([complementary_noisy_input_ids, original_input_ids], dim=1)
                    embeds_pair2 = torch.cat([complementary_noisy_embeds, original_embeds], dim=1)
                    labels_pair2 = complementary_labels
                    position_ids_pair2 = original_position_ids
                
                    # Batch 拼接
                    input_ids = torch.cat([input_ids_pair1, input_ids_pair2], dim=0)
                    inputs_embeds = torch.cat([embeds_pair1, embeds_pair2], dim=0)
                    labels = torch.cat([labels_pair1, labels_pair2], dim=0)
                    position_ids = torch.cat([position_ids_pair1, position_ids_pair2], dim=1)
                
                if _debug_fwd:
                    torch.cuda.synchronize()
                    _t_preM = _time.perf_counter()

                if self.deep_json_scaffold and self._section_tokenizer is not None:
                    # Deep scaffold v2: always uses hybrid block causal mask
                    attention_mask = self.gen_hybrid_block_causal_mask(L, response_block_idx, turn_idx, input_ids.shape[0], self.config.num_attention_heads)
                elif self.use_block_causal_mask:
                    if self.block_causal_no_dynamic:
                        attention_mask = self.gen_block_causal_mask(L, bd_size, input_ids.shape[0], self.config.num_attention_heads)
                    else:
                        attention_mask = self.gen_hybrid_block_causal_mask(L, response_block_idx, turn_idx, input_ids.shape[0], self.config.num_attention_heads)
                else:
                    attention_mask = self.gen_mask(L, bd_size, input_ids.shape[0], self.config.num_attention_heads)

                if _debug_fwd:
                    torch.cuda.synchronize()
                    _t_postM = _time.perf_counter()

                # 清空 pixel_values（已替换）
                pixel_values = None
                pixel_values_videos = None


            # Phase D: CP sharding — shard doubled sequences before model call
            if is_cp_enabled():
                cp_sz = get_cp_size()
                cp_rk = get_cp_rank()

                if input_ids is not None:
                    input_ids = shard_doubled_sequence(input_ids, seq_dim=1)
                if inputs_embeds is not None:
                    inputs_embeds = shard_doubled_sequence(inputs_embeds, seq_dim=1)

                if position_ids is not None:
                    position_ids = shard_position_ids(position_ids)
                else:
                    # Text-only path: generate explicit CP-aware position_ids.
                    # Each rank owns [rank*L/cp .. (rank+1)*L/cp) positions,
                    # shared by both x_t and x_0 halves (the inner model's
                    # auto-generated ids would start from 0 on every rank).
                    local_half = seq_len // cp_sz
                    pos_start = cp_rk * local_half
                    pos = torch.arange(
                        pos_start, pos_start + local_half,
                        device=input_ids.device if input_ids is not None else inputs_embeds.device,
                    )
                    B_eff = (input_ids.shape[0] if input_ids is not None
                             else inputs_embeds.shape[0])
                    # (3, B, local_half)
                    position_ids = pos.view(1, 1, -1).expand(3, B_eff, -1)

            # Section-MoE-LoRA: set section_ids before language model forward
            if self._use_section_moe_lora and _block_to_section is not None:
                _sec_ids = _build_section_ids_tensor(
                    response_block_idx, _block_to_section, seq_len,
                    device=input_ids.device if input_ids is not None else inputs_embeds.device,
                )
                # Doubled sequence: [noisy | clean] share the same section layout
                _sec_ids_doubled = _sec_ids.repeat(2)               # [2*seq_len]
                B_eff = (input_ids.shape[0] if input_ids is not None
                         else inputs_embeds.shape[0])
                _sec_ids_batch = _sec_ids_doubled.unsqueeze(0).expand(B_eff, -1)
                if is_cp_enabled():
                    _sec_ids_batch = shard_doubled_sequence(_sec_ids_batch, seq_dim=1)
                _set_section_ids(_sec_ids_batch)

            # Phase D: 调用内层（多模态时传 inputs_embeds，纯文本时传 input_ids）
            if _debug_fwd:
                torch.cuda.synchronize()
                _t_preD = _time.perf_counter()

            if pixel_values is None and pixel_values_videos is None:
                # 纯文本：传 input_ids（内层会 embed）
                outputs = self.model(
                    input_ids=input_ids,
                    pixel_values=None,
                    pixel_values_videos=None,
                    image_grid_thw=None,
                    video_grid_thw=None,
                    position_ids=position_ids,
                    attention_mask=attention_mask,
                    past_key_values=past_key_values,
                    inputs_embeds=inputs_embeds,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                    output_hidden_states=output_hidden_states,
                    return_dict=True,
                    cache_position=cache_position,
                    update_kv_cache=update_kv_cache,
                    bd_size=bd_size,
                    **kwargs,
                )
            else:
                # 多模态：传 inputs_embeds（已 masked_scatter）
                outputs = self.model.language_model(
                    input_ids=None,
                    position_ids=position_ids,
                    attention_mask=attention_mask,
                    past_key_values=past_key_values,
                    inputs_embeds=inputs_embeds,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                    output_hidden_states=output_hidden_states,
                    return_dict=True,
                    cache_position=cache_position,
                    update_kv_cache=update_kv_cache,
                    bd_size=bd_size,
                    **kwargs,
                )

            if _debug_fwd:
                torch.cuda.synchronize()
                _t_postD = _time.perf_counter()
                _rank = int(_os.environ.get("RANK", "0"))
                if _rank == 0:
                    print(f"[FWD_TIMING] seq={L} phaseAB={_t_phaseAB-_t_phaseA:.4f} "
                          f"phaseC_prep={_t_preM-_t_phaseAB:.4f} "
                          f"create_mask={_t_postM-_t_preM:.4f} "
                          f"inner_model={_t_postD-_t_preD:.4f} "
                          f"total_fwd={_t_postD-_t_phaseA:.4f}", flush=True)

        else:
            outputs = self.model(
                    input_ids=input_ids,
                    pixel_values=pixel_values,
                    pixel_values_videos=pixel_values_videos,
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    position_ids=position_ids,
                    attention_mask=attention_mask,
                    past_key_values=past_key_values,
                    inputs_embeds=inputs_embeds,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                    output_hidden_states=output_hidden_states,
                    return_dict=True,
                    cache_position=cache_position,
                    update_kv_cache=update_kv_cache,
                    bd_size=eval_bd_size,
                    **kwargs,
                )
                
        hidden_states = outputs[0]
        loss = None
        if self.training:
            mdm_hidden_states = hidden_states[:, :hidden_states.shape[1]//2, :]
            # 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
            logits = self.lm_head(mdm_hidden_states[:, slice_indices, :])

            # CP: shard labels to match local logits
            if is_cp_enabled():
                cp_sz = get_cp_size()
                cp_rk = get_cp_rank()
                local_label_len = labels.shape[1] // cp_sz
                cp_label_start = cp_rk * local_label_len
                labels = labels[:, cp_label_start:cp_label_start + local_label_len].contiguous()

            if self.use_block_causal_mask:
                new_kwargs = {
                    'num_items_in_batch': 2*kwargs['num_items_in_batch'],
                }
            else:
                new_kwargs = kwargs

            # SASD: section-importance-weighted loss
            _use_section_weighted_loss = (
                self.section_loss_weights is not None
                and self.deep_json_scaffold
                and self._section_tokenizer is not None
            )

            if labels is not None:
                if _use_section_weighted_loss:
                    _sasd_weight = self._build_section_weight_tensor(labels, response_block_idx, n_blocks, block_to_section=_block_to_section)
                    # CP: shard weight tensor to match local labels
                    if is_cp_enabled():
                        _sasd_weight = _sasd_weight[:, cp_label_start:cp_label_start + local_label_len].contiguous()
                    loss = self.compute_section_weighted_loss(
                        logits, labels, _sasd_weight,
                        self.config.text_config.vocab_size,
                        num_items_in_batch=new_kwargs.get('num_items_in_batch', None),
                    )
                else:
                    loss = self.loss_function(
                        logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **new_kwargs
                    )
            if self.use_block_causal_mask:
                # CP: shard original_labels to match causal logits
                if is_cp_enabled():
                    original_labels = original_labels[:, cp_label_start:cp_label_start + local_label_len].contiguous()

                if self.complementary_mask:
                    causal_hidden_states = hidden_states[:hidden_states.shape[0]//2, hidden_states.shape[1]//2:, :]
                else:
                    causal_hidden_states = hidden_states[:, :hidden_states.shape[1]//2, :]
                causal_logits = self.lm_head(causal_hidden_states[:, slice_indices, :])
                loss += self.loss_function(
                    logits=causal_logits, labels=original_labels, vocab_size=self.config.text_config.vocab_size, **new_kwargs
                )

            if self.entropy_loss:
                # Use num_items_in_batch for global normalization (consistent with cross entropy)
                num_items = kwargs.get('num_items_in_batch', None)
                entropy_loss = self.compute_entropy_loss(logits, labels, num_items_in_batch=num_items)
                loss += self.entropy_loss_weight * entropy_loss

            # CP: scale local loss so FSDP gradient sync produces correct results.
            # Each CP rank computes loss on 1/cp_size of the tokens, but
            # num_items_in_batch (from the Trainer) still reflects the FULL
            # sample's token count.  This means loss_local ≈ loss_full / cp_size,
            # and the resulting gradient is also scaled by 1/cp_size.
            # FSDP averages gradients across ALL ranks (including CP peers),
            # so without correction the effective lr would be lr/cp_size.
            # Multiplying by cp_size restores the correct gradient magnitude.
            if is_cp_enabled() and loss is not None:
                loss = loss * get_cp_size()
        else:
            slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
            logits = self.lm_head(hidden_states[:, slice_indices, :])
        
        return Fast_dDriveCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=outputs.rope_deltas,
        )

    @torch.no_grad()
    def sbd_inference_quadratic(
        self,
        clean_input_ids: Optional[torch.Tensor],
        draft_input_ids: torch.Tensor,
        block_length: int,
        draft_only: bool = False,
        past_key_values: Optional[Cache] = None,
        use_cache: bool = False,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        mask_token_id: Optional[int] = None,
    ):
        """Quadratic speculative decoding: one forward for full block (draft_only or full quadratic)."""
        text_config = self.config.text_config
        prev_mode = getattr(text_config, "self_spec_inference_mode", None)
        prev_block_length = getattr(text_config, "block_length", None)
        mask_id = mask_token_id if mask_token_id is not None else getattr(text_config, "mask_token_id", 151665)

        try:
            text_config.self_spec_inference_mode = "default" if draft_only else "quadratic"
            text_config.block_length = block_length

            if draft_only:
                assert clean_input_ids is not None
                if use_cache and past_key_values is None:
                    past_key_values = DynamicCache(config=text_config)
                input_ids = torch.cat([clean_input_ids, draft_input_ids], dim=-1)
                output = self.forward(
                    input_ids=input_ids,
                    pixel_values=pixel_values,
                    image_grid_thw=image_grid_thw,
                    past_key_values=past_key_values,
                    use_cache=use_cache,
                    update_kv_cache=True,
                )
                logits = output.logits
                past_key_values = output.past_key_values
                if use_cache and past_key_values is not None:
                    past_key_values = _crop_dynamic_cache(past_key_values, clean_input_ids.shape[1])
                return logits, past_key_values

            draft_len = block_length * (block_length + 1)
            draft_input_ids = torch.cat(
                [
                    draft_input_ids.view(-1, block_length, 1),
                    torch.full(
                        (draft_input_ids.shape[0], block_length, block_length),
                        fill_value=mask_id,
                        device=draft_input_ids.device,
                    ),
                ],
                dim=-1,
            ).view(-1, draft_len)

            if use_cache:
                assert past_key_values is not None
                assert clean_input_ids is None
                clean_len = past_key_values.get_seq_length()
                input_ids = draft_input_ids
            else:
                clean_len = clean_input_ids.shape[1]
                input_ids = torch.cat([clean_input_ids, draft_input_ids], dim=-1)

            # Account for rope_deltas (image tokens compress positions in Qwen VL)
            rope_deltas = getattr(self.model, "rope_deltas", None)
            if rope_deltas is not None:
                pos_offset = int((clean_len + rope_deltas).squeeze().item())
            else:
                pos_offset = clean_len

            # Build per-block position ids for the quadratic structure:
            # Group i (0..block_length-1) has positions [pos_offset+i, pos_offset+i+1, ..., pos_offset+i+block_length]
            per_block_position_ids = torch.arange(
                pos_offset, pos_offset + block_length + 1, device=draft_input_ids.device
            )[None,].repeat(block_length, 1)
            per_block_position_ids = per_block_position_ids + torch.arange(
                block_length, device=draft_input_ids.device
            ).view(-1, 1)

            if use_cache:
                position_ids_1d = per_block_position_ids.view(-1)
            else:
                clean_position_ids = torch.arange(pos_offset - clean_len, pos_offset, device=draft_input_ids.device)
                position_ids_1d = torch.cat([clean_position_ids, per_block_position_ids.view(-1)], dim=-1)

            # Expand to 3D: (3, batch, seq_len) — all 3 dims same for text tokens
            batch_size = input_ids.shape[0]
            position_ids = position_ids_1d.view(1, 1, -1).expand(3, batch_size, -1)

            output = self.forward(
                input_ids=input_ids,
                pixel_values=pixel_values if not use_cache else None,
                image_grid_thw=image_grid_thw if not use_cache else None,
                past_key_values=past_key_values,
                use_cache=use_cache,
                position_ids=position_ids,
                update_kv_cache=True,
            )
            logits = output.logits
            past_key_values = output.past_key_values
            if use_cache and past_key_values is not None:
                past_key_values = _extract_draft_kv_cache(past_key_values, clean_len, block_length)
            return logits, past_key_values
        finally:
            text_config.self_spec_inference_mode = prev_mode
            text_config.block_length = prev_block_length

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

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
            position_ids=position_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            second_per_grid_ts=second_per_grid_ts,
            use_cache=use_cache,
            **kwargs,
        )

        # Qwen2-5-VL position_ids are prepared with rope_deltas
        if position_ids is None:
            # Calculate RoPE index once per generation in the pre-fill stage only.
            # When compiling, we can't check tensor values thus we check only input length
            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
            # models currently cannot do assisted decoding
            if cache_position[0] == 0 or self.model.rope_deltas is None:
                vision_positions, rope_deltas = self.model.get_rope_index(
                    model_inputs.get("input_ids", None),
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    attention_mask=attention_mask,
                )
                self.model.rope_deltas = rope_deltas
            # then use the prev pre-calculated rope-deltas to get the correct position ids
            elif "position_ids" in model_inputs:
                batch_size, seq_length = model_inputs["position_ids"].shape
                device = model_inputs["position_ids"].device
                position_ids = torch.arange(seq_length, device=device)
                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
                delta = cache_position[0] + self.model.rope_deltas
                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                vision_positions = position_ids + delta.expand_as(position_ids)

            # Concatenate "text + vision" positions into [4, bs, seq-len]
            text_positions = model_inputs["position_ids"][None, ...]
            model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)

        if cache_position[0] != 0:
            model_inputs["pixel_values"] = None
            model_inputs["pixel_values_videos"] = None

        return model_inputs

    def _get_image_nums_and_video_nums(
        self,
        input_ids: Optional[torch.LongTensor],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
        These parameters are not passed through the processor to avoid unpredictable impacts from interface modifications.

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary.

        Returns:
            image_nums (`torch.LongTensor` of shape `(batch_size, num_images_sample)`)
            video_nums (`torch.LongTensor` of shape `(batch_size, num_videos_sample)`)
        """
        image_token_id = self.config.image_token_id
        video_token_id = self.config.video_token_id
        vision_start_token_id = self.config.vision_start_token_id

        if inputs_embeds is not None:
            vision_start_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(vision_start_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
            image_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(image_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
            video_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(video_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
        else:
            vision_start_mask = input_ids == vision_start_token_id
            image_mask = input_ids == image_token_id
            video_mask = input_ids == video_token_id

        vision_first_mask = torch.roll(vision_start_mask, shifts=1, dims=1)
        image_nums = torch.sum(vision_first_mask & image_mask, dim=1)
        video_nums = torch.sum(vision_first_mask & video_mask, dim=1)

        return image_nums, video_nums

    def _expand_inputs_for_generation(
        self,
        expand_size: int = 1,
        is_encoder_decoder: bool = False,
        input_ids: Optional[torch.LongTensor] = None,
        **model_kwargs,
    ) -> tuple[torch.LongTensor, dict[str, Any]]:
        # Overwritten -- Support for expanding tensors without a batch size dimension
        # e.g., pixel_values, image_grid_thw, pixel_values_videos, video_grid_thw, second_per_grid_t
        # pixel_values.shape[0] is sum(seqlen_images for samples)
        # image_grid_thw.shape[0] is sum(num_images for samples)

        if expand_size == 1:
            return input_ids, model_kwargs

        visual_keys = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw", "second_per_grid_ts"]

        def _expand_dict_for_generation_visual(dict_to_expand):
            image_grid_thw = model_kwargs.get("image_grid_thw", None)
            video_grid_thw = model_kwargs.get("video_grid_thw", None)
            image_nums, video_nums = self._get_image_nums_and_video_nums(
                input_ids, inputs_embeds=model_kwargs.get("inputs_embeds", None)
            )

            def _repeat_interleave_samples(x, lengths, repeat_times):
                samples = torch.split(x, lengths)
                repeat_args = [repeat_times] + [1] * (x.dim() - 1)
                result = torch.cat([sample.repeat(*repeat_args) for sample in samples], dim=0)
                return result

            for key in dict_to_expand:
                if key == "pixel_values":
                    # split images into samples
                    samples = torch.split(image_grid_thw, list(image_nums))
                    # compute the sequence length of images for each sample
                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "image_grid_thw":
                    # get the num of images for each sample
                    lengths = list(image_nums)
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "pixel_values_videos":
                    samples = torch.split(video_grid_thw, list(video_nums))
                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "video_grid_thw":
                    lengths = list(video_nums)
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "second_per_grid_ts":
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=list(video_nums), repeat_times=expand_size
                    )
            return dict_to_expand

        def _expand_dict_for_generation(dict_to_expand):
            for key in dict_to_expand:
                if (
                    key != "cache_position"
                    and dict_to_expand[key] is not None
                    and isinstance(dict_to_expand[key], torch.Tensor)
                    and key not in visual_keys
                ):
                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
            return dict_to_expand

        model_kwargs = _expand_dict_for_generation_visual(model_kwargs)

        if input_ids is not None:
            input_ids = input_ids.repeat_interleave(expand_size, dim=0)

        model_kwargs = _expand_dict_for_generation(model_kwargs)

        if is_encoder_decoder:
            if model_kwargs.get("encoder_outputs") is None:
                raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
            model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])

        return input_ids, model_kwargs


# Attach inference-time decoding methods (Section Diffusion, Scaffold Spec,
# Scaffold Spec with multi-trajectory rollouts) onto the model class.  These
# live in ``generation_utils.py`` to keep this file focused on the model
# definition itself.  The explicit ``from .generation_utils import …`` form is
# required so HF's ``trust_remote_code`` loader picks up the dependency.
from .generation_utils import attach_generation_methods as _attach_generation_methods
_attach_generation_methods(Fast_dDriveForConditionalGeneration)


__all__ = ["Fast_dDriveForConditionalGeneration", "Fast_dDriveModel", "Fast_dDrivePreTrainedModel", "Fast_dDriveTextModel"]