diff --git "a/modeling.py" "b/modeling.py"
new file mode 100644--- /dev/null
+++ "b/modeling.py"
@@ -0,0 +1,2558 @@
+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_dVLMConfig, Fast_dVLMTextConfig, Fast_dVLMVisionConfig
+
+from torch.nn.attention.flex_attention import flex_attention, create_block_mask
+
+from functools import partial
+import random
+import math
+
+logger = logging.get_logger(__name__)
+
+
+# @torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
+# @torch.compile()
+def fused_flex_attention(q, k, v, mask=None):
+    return flex_attention(q, k, v, block_mask=mask, enable_gqa=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
+
+class Fast_dVLMMLP(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_dVLMVisionPatchEmbed(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_dVLMVisionRotaryEmbedding(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_dVLMPatchMerger(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_dVLMVisionAttention(nn.Module):
+    def __init__(self, config: Fast_dVLMVisionConfig) -> 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_dVLMVisionBlock(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_dVLMVisionAttention(config=config)
+        self.mlp = Fast_dVLMMLP(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_dVLMPreTrainedModel(PreTrainedModel):
+    config: Fast_dVLMConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Fast_dVLMDecoderLayer", "Fast_dVLMVisionBlock"]
+    _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_dVLMVisionTransformerPretrainedModel(Fast_dVLMPreTrainedModel):
+    config: Fast_dVLMVisionConfig
+    _no_split_modules = ["Fast_dVLMVisionBlock"]
+
+    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_dVLMVisionPatchEmbed(
+            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_dVLMVisionRotaryEmbedding(head_dim // 2)
+
+        self.blocks = nn.ModuleList([Fast_dVLMVisionBlock(config) for _ in range(config.depth)])
+        self.merger = Fast_dVLMPatchMerger(
+            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_dVLMModelOutputWithPast(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_dVLMRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Fast_dVLMTextConfig, 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_dVLM 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_dVLMAttention(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_dVLMTextConfig, 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_dVLMRotaryEmbedding(config=config)
+
+    @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)
+
+        if self.training:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+            
+            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_dVLMDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Fast_dVLMTextConfig, 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_dVLMAttention(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_dVLMTextModel(Fast_dVLMPreTrainedModel):
+    config: Fast_dVLMTextConfig
+
+    def __init__(self, config: Fast_dVLMTextConfig):
+        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_dVLMDecoderLayer(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_dVLMRotaryEmbedding(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]:
+        r"""
+        update_kv_cache (`bool`, *optional*, defaults to `False`):
+            Whether to update the KV cache with the current forward pass outputs.
+        """
+        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_dVLMModel(Fast_dVLMPreTrainedModel):
+    base_model_prefix = ""
+    _checkpoint_conversion_mapping = {"^model": "language_model"}
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
+    config: Fast_dVLMConfig
+    _no_split_modules = ["Fast_dVLMDecoderLayer", "Fast_dVLMVisionBlock"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.visual = Fast_dVLMVisionTransformerPretrainedModel._from_config(config.vision_config)
+        self.language_model = Fast_dVLMTextModel._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
+
+    @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_dVLMModelOutputWithPast]:
+        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.
+        update_kv_cache (`bool`, *optional*, defaults to `False`):
+            Whether to update the KV cache with the current forward pass outputs.
+        bd_size (`int`, *optional*):
+            Block diffusion size to use for this forward pass. Overrides the model default when set.
+        """
+        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)
+        elif self.training and pixel_values_videos is None:
+            # Dummy forward through vision encoder to keep FSDP/DeepSpeed gradient
+            # sync alive when this rank has a text-only batch but other ranks have images.
+            # Mirrors VILA's __embed_media_tokens dummy trick (llava_arch.py L519-526).
+            _merge = self.visual.spatial_merge_size  # typically 2
+            _t = self.visual.patch_embed.temporal_patch_size
+            _p = self.visual.patch_embed.patch_size
+            _c = self.visual.patch_embed.in_channels  # 3
+            # grid_thw [1, _merge, _merge] → seq_len = _merge*_merge = spatial_merge_unit
+            dummy_pixel = torch.zeros(
+                _merge * _merge, _t * _p * _p * _c,
+                dtype=inputs_embeds.dtype, device=inputs_embeds.device,
+            )
+            dummy_grid = torch.tensor([[1, _merge, _merge]], dtype=torch.long, device=inputs_embeds.device)
+            dummy_embeds = self.visual(dummy_pixel, grid_thw=dummy_grid)
+            inputs_embeds = inputs_embeds + dummy_embeds.sum() * 0
+
+        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:
+            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)
+
+        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_dVLMModelOutputWithPast(
+            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_dVLM causal language model (or autoregressive) outputs.
+    """
+)
+class Fast_dVLMCausalLMOutputWithPast(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_dVLMForConditionalGeneration(Fast_dVLMPreTrainedModel, GenerationMixin):
+    config_class = Fast_dVLMConfig
+    _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
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Fast_dVLMModel(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.im_end_token_id = 151645  # <|im_end|> token id
+        # 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 = create_block_mask(
+            # 这里将原来的 block_size=block_size 改为传入 Tensor
+            partial(block_diff_mask, block_size=block_size_t, n=n_t),
+            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 = create_block_mask(
+            partial(block_causal_mask, block_size=block_size_t, n=n_t),
+            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 = create_block_mask(
+            partial(hybrid_block_causal_mask_multiturn, response_block_idx=response_block_idx, turn_idx=turn_idx, n=n_t),
+            B=B, H=H, Q_LEN=seqlen*2, KV_LEN=seqlen*2
+        )
+        return mask
+
+    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_dVLMCausalLMOutputWithPast]:
+        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.
+        mask_id (`int`, *optional*, defaults to `151665`):
+            Token id used as the mask placeholder for block diffusion.
+        update_kv_cache (`bool`, *optional*, defaults to `False`):
+            Whether to update the KV cache with the current forward pass outputs.
+        eval_bd_size (`int`, *optional*):
+            Block diffusion size to use during evaluation. Overrides the model default when set.
+        """
+        # 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
+                # sqrt mapping: larger block sizes get more training time
+                scaled_ratio = math.sqrt(update_ratio)
+                idx = min(int(scaled_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
+            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
+
+                if 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
+                    
+                    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.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
+
+                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)
+                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_block_idx, turn_idx, n_blocks = self.compute_response_block_idx(labels, bd_size)
+                # 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
+
+                if 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)
+
+                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
+                    
+                    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 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)
+                
+                # 清空 pixel_values（已替换）
+                pixel_values = None
+                pixel_values_videos = None
+
+
+            # Phase D: 调用内层（多模态时传 inputs_embeds，纯文本时传 input_ids）
+            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,
+                )
+
+        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, :])
+
+            if self.use_block_causal_mask:
+                new_kwargs = {
+                    'num_items_in_batch': 2*kwargs['num_items_in_batch'],
+                }
+            else:
+                new_kwargs = kwargs
+            if labels is not None:
+                loss = self.loss_function(
+                    logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **new_kwargs
+                ) * 0.5
+            if self.use_block_causal_mask:
+                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
+        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_dVLMCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            rope_deltas=outputs.rope_deltas,
+        )
+
+    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
+
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids,
+        tokenizer,
+        block_size=32,
+        max_tokens=1024,
+        pixel_values=None,
+        image_grid_thw=None,
+        mask_id=151665,
+        stop_token=151645,
+    ):
+        """Speculative block-causal parallel decoding for Fast-dVLM.
+
+        Each iteration: (1) draft a block of masked tokens via one diffusion
+        forward pass — all masked positions are filled at once, (2) verify
+        with an AR forward pass and accept the longest matching prefix, then
+        cache the accepted tokens.
+
+        Args:
+            input_ids: Prompt token ids ``[1, prompt_len]``.
+            tokenizer: Tokenizer (reserved for future use).
+            block_size: Number of tokens to draft per block.
+            max_tokens: Maximum new tokens to generate.
+            pixel_values: Optional image pixel values for VLM.
+            image_grid_thw: Optional image grid info for VLM.
+            mask_id: Token id used as the ``[MASK]`` placeholder.
+            stop_token: EOS token id for early stopping.
+
+        Returns:
+            Generated token ids ``[1, prompt_len + gen_len]``.
+        """
+        self.model.bd_size = block_size
+        original_input_length = input_ids.shape[1]
+
+        # Prefill: encode prompt and generate first token autoregressively
+        output = self.forward(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            image_grid_thw=image_grid_thw,
+            use_cache=True,
+            update_kv_cache=True,
+        )
+        logits, past_key_values = output.logits, output.past_key_values
+        next_token = logits[:, -1:, :].argmax(dim=-1)
+        input_ids = torch.cat([input_ids, next_token], dim=1)
+
+        while True:
+            prompt_length = input_ids.shape[1]
+
+            # Append (block_size - 1) mask tokens as draft placeholders
+            x_init = mask_id * torch.ones(
+                (input_ids.shape[0], block_size - 1), device=self.device, dtype=torch.long
+            )
+            x_t = torch.cat([input_ids, x_init], dim=1)
+
+            # Draft: diffusion forward fills all masked positions in one pass
+            current_ids = x_t[:, -block_size:]
+            logits = self.forward(
+                input_ids=current_ids,
+                use_cache=True,
+                past_key_values=past_key_values,
+                update_kv_cache=False,
+                eval_bd_size=block_size,
+            ).logits
+
+            logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
+            x_1 = logits.argmax(dim=-1)
+            mask_idx = current_ids == mask_id
+            x_t[:, -block_size:][mask_idx] = x_1[mask_idx]
+
+            # Verify: AR forward to accept longest correct prefix
+            current_ids = x_t[:, -block_size:]
+            output = self.forward(
+                input_ids=current_ids,
+                use_cache=True,
+                past_key_values=past_key_values,
+                update_kv_cache=True,
+                eval_bd_size=block_size,
+            )
+            logits, past_key_values = output.logits, output.past_key_values
+            ar_block_token = logits.argmax(dim=-1)
+
+            accepted_token_num = 1
+            for i in range(ar_block_token.shape[1] - 1):
+                if (ar_block_token[:, i] == current_ids[:, i + 1]).all():
+                    accepted_token_num += 1
+                else:
+                    break
+
+            input_ids = torch.cat([input_ids, ar_block_token[:, :accepted_token_num]], dim=1)
+
+            # Crop KV cache to match accepted length
+            new_past_key_values = []
+            for layer_num in range(len(past_key_values)):
+                layer_past_key_values = ()
+                for kv_idx in range(len(past_key_values[layer_num])):
+                    layer_past_key_values += (
+                        past_key_values[layer_num][kv_idx][:, :, : input_ids.shape[1] - 1, :],
+                    )
+                new_past_key_values.append(layer_past_key_values)
+            past_key_values = DynamicCache(new_past_key_values)
+
+            if input_ids.shape[1] - original_input_length > max_tokens:
+                break
+            if stop_token in input_ids[:, prompt_length:]:
+                stop_token_idx = (input_ids[:, prompt_length:] == stop_token).nonzero()[0][1]
+                if (input_ids[:, prompt_length : prompt_length + stop_token_idx] == mask_id).sum() == 0:
+                    break
+
+        # Truncate at stop token
+        if stop_token in input_ids[:, original_input_length:]:
+            stop_token_idx = (input_ids[:, original_input_length:] == stop_token).nonzero()[0][1]
+            input_ids = input_ids[:, : stop_token_idx + original_input_length + 1]
+
+        return input_ids
+
+
+__all__ = ["Fast_dVLMForConditionalGeneration", "Fast_dVLMModel", "Fast_dVLMPreTrainedModel", "Fast_dVLMTextModel"]
\ No newline at end of file