modeling_smb_vision.py

# coding=utf-8
# Copyright 2025 The SMB Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from dataclasses import dataclass
from typing import Callable, Optional

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

from transformers.activations import ACT2FN
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutput, ModelOutput
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs
from .configuration_smb_vision import (
    SMBVisionConfig,
    SMBVisionPredictorConfig,
    SMBVisionModelConfig,
)


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 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: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

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

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

    return attn_output, attn_weights


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

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


class SMBVisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.linear_fc1 = nn.Linear(self.hidden_size, self.intermediate_size, bias=True)
        self.linear_fc2 = nn.Linear(self.intermediate_size, self.hidden_size, bias=True)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_state):
        return self.linear_fc2(self.act_fn(self.linear_fc1(hidden_state)))


class SMBVisionPatchEmbed(nn.Module):
    def __init__(self, config) -> None:
        super().__init__()
        self.patch_size = config.patch_size
        self.temporal_patch_size = config.temporal_patch_size
        self.in_channels = config.in_channels
        self.embed_dim = config.hidden_size

        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
        for in_channels in [1, 3, 4]:
            setattr(
                self,
                f"proj_c{in_channels}",
                nn.Conv3d(
                    in_channels,
                    self.embed_dim,
                    kernel_size=kernel_size,
                    stride=kernel_size,
                    bias=True,
                ),
            )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj_c1.weight.dtype
        if self.in_channels == 1:  # grayscale
            hidden_states = hidden_states.view(
                -1, 1, self.temporal_patch_size, self.patch_size, self.patch_size
            )
            hidden_states = self.proj_c1(hidden_states.to(dtype=target_dtype)).view(
                -1, self.embed_dim
            )
        elif self.in_channels == 3:  # rgb
            hidden_states = hidden_states.view(
                -1, 3, self.temporal_patch_size, self.patch_size, self.patch_size
            )
            hidden_states = self.proj_c3(hidden_states.to(dtype=target_dtype)).view(
                -1, self.embed_dim
            )
        elif self.in_channels == 4:  # multi sequence
            hidden_states = hidden_states.view(
                -1, 4, self.temporal_patch_size, self.patch_size, self.patch_size
            )
            hidden_states = self.proj_c4(hidden_states.to(dtype=target_dtype)).view(
                -1, self.embed_dim
            )
        else:
            raise ValueError(f"Unsupported number of channels: {self.in_channels}")
        return hidden_states


class SMBVisionRotaryEmbedding(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 SMBVisionPatchMerger(nn.Module):
    def __init__(self, config, use_postshuffle_norm=False) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size * (config.spatial_merge_size**2)
        self.use_postshuffle_norm = use_postshuffle_norm
        self.norm = nn.LayerNorm(
            self.hidden_size if use_postshuffle_norm else config.hidden_size, eps=1e-6
        )
        self.linear_fc1 = nn.Linear(self.hidden_size, self.hidden_size)
        self.act_fn = nn.GELU()
        self.linear_fc2 = nn.Linear(self.hidden_size, config.out_hidden_size)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.norm(
            x.contiguous().view(-1, self.hidden_size)
            if self.use_postshuffle_norm
            else x
        ).view(-1, self.hidden_size)
        x = self.linear_fc2(self.act_fn(self.linear_fc1(x)))
        return x


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


class SMBVisionAttention(nn.Module):
    def __init__(self, config) -> 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 SMBVisionBlock(GradientCheckpointingLayer):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = nn.LayerNorm(config.hidden_size, eps=1e-6)
        self.norm2 = nn.LayerNorm(config.hidden_size, eps=1e-6)
        self.attn = SMBVisionAttention(config=config)
        self.mlp = SMBVisionMLP(config=config)

    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


class SMBVisionEncoder(PreTrainedModel):
    config: SMBVisionConfig
    _no_split_modules = ["SMBVisionBlock"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
    _supports_attention_backend = True

    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.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size

        self.patch_embed = SMBVisionPatchEmbed(
            config=config,
        )

        self.pos_embed = nn.Embedding(
            config.num_position_embeddings, config.hidden_size
        )
        self.num_grid_per_side = int(config.num_position_embeddings**0.5)

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

        self.blocks = nn.ModuleList(
            [SMBVisionBlock(config) for _ in range(config.depth)]
        )
        self.merger = SMBVisionPatchMerger(
            config=config,
            use_postshuffle_norm=False,
        )

        self.deepstack_visual_indexes = config.deepstack_visual_indexes
        self.deepstack_merger_list = nn.ModuleList(
            [
                SMBVisionPatchMerger(
                    config=config,
                    use_postshuffle_norm=True,
                )
                for _ in range(len(config.deepstack_visual_indexes))
            ]
        )

        self.gradient_checkpointing = False

    def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
        merge_size = self.spatial_merge_size

        max_hw = int(grid_thw[:, 1:].max().item())
        freq_table = self.rotary_pos_emb(max_hw)  # (max_hw, dim // 2)
        device = freq_table.device

        total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
        pos_ids = torch.empty((total_tokens, 2), dtype=torch.long, device=device)

        offset = 0
        for num_frames, height, width in grid_thw:
            merged_h, merged_w = height // merge_size, width // merge_size

            block_rows = torch.arange(merged_h, device=device)  # block row indices
            block_cols = torch.arange(merged_w, device=device)  # block col indices
            intra_row = torch.arange(
                merge_size, device=device
            )  # intra-block row offsets
            intra_col = torch.arange(
                merge_size, device=device
            )  # intra-block col offsets

            # Compute full-resolution positions
            row_idx = (
                block_rows[:, None, None, None] * merge_size
                + intra_row[None, None, :, None]
            )
            col_idx = (
                block_cols[None, :, None, None] * merge_size
                + intra_col[None, None, None, :]
            )

            row_idx = row_idx.expand(
                merged_h, merged_w, merge_size, merge_size
            ).reshape(-1)
            col_idx = col_idx.expand(
                merged_h, merged_w, merge_size, merge_size
            ).reshape(-1)

            coords = torch.stack((row_idx, col_idx), dim=-1)

            if num_frames > 1:
                coords = coords.repeat(num_frames, 1)

            num_tokens = coords.shape[0]
            pos_ids[offset : offset + num_tokens] = coords
            offset += num_tokens

        embeddings = freq_table[pos_ids]  # lookup rotary embeddings
        embeddings = embeddings.flatten(1)
        return embeddings

    def fast_pos_embed_interpolate(self, grid_thw):
        grid_ts, grid_hs, grid_ws = grid_thw[:, 0], grid_thw[:, 1], grid_thw[:, 2]

        idx_list = [[] for _ in range(4)]
        weight_list = [[] for _ in range(4)]

        for t, h, w in zip(grid_ts, grid_hs, grid_ws):
            h_idxs = torch.linspace(0, self.num_grid_per_side - 1, h)
            w_idxs = torch.linspace(0, self.num_grid_per_side - 1, w)

            h_idxs_floor = h_idxs.int()
            w_idxs_floor = w_idxs.int()
            h_idxs_ceil = (h_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)
            w_idxs_ceil = (w_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)

            dh = h_idxs - h_idxs_floor
            dw = w_idxs - w_idxs_floor

            base_h = h_idxs_floor * self.num_grid_per_side
            base_h_ceil = h_idxs_ceil * self.num_grid_per_side

            indices = [
                (base_h[None].T + w_idxs_floor[None]).flatten(),
                (base_h[None].T + w_idxs_ceil[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_floor[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_ceil[None]).flatten(),
            ]

            weights = [
                ((1 - dh)[None].T * (1 - dw)[None]).flatten(),
                ((1 - dh)[None].T * dw[None]).flatten(),
                (dh[None].T * (1 - dw)[None]).flatten(),
                (dh[None].T * dw[None]).flatten(),
            ]

            for i in range(4):
                idx_list[i].extend(indices[i].tolist())
                weight_list[i].extend(weights[i].tolist())

        idx_tensor = torch.tensor(idx_list, dtype=torch.long, device=self.pos_embed.weight.device)
        weight_tensor = torch.tensor(
            weight_list, dtype=self.pos_embed.weight.dtype, device=self.pos_embed.weight.device
        )
        pos_embeds = self.pos_embed(idx_tensor) * weight_tensor[:, :, None]
        patch_pos_embeds = pos_embeds[0] + pos_embeds[1] + pos_embeds[2] + pos_embeds[3]

        patch_pos_embeds = patch_pos_embeds.split([h * w for h, w in zip(grid_hs, grid_ws)])

        patch_pos_embeds_permute = []
        merge_size = self.config.spatial_merge_size
        for pos_embed, t, h, w in zip(patch_pos_embeds, grid_ts, grid_hs, grid_ws):
            pos_embed = pos_embed.repeat(t, 1)
            pos_embed = (
                pos_embed.view(t, h // merge_size, merge_size, w // merge_size, merge_size, -1)
                .permute(0, 1, 3, 2, 4, 5)
                .flatten(0, 4)
            )
            patch_pos_embeds_permute.append(pos_embed)
        patch_pos_embeds = torch.cat(patch_pos_embeds_permute)
        return patch_pos_embeds

    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)

        pos_embeds = self.fast_pos_embed_interpolate(grid_thw)
        hidden_states = hidden_states + pos_embeds

        rotary_pos_emb = self.rot_pos_emb(grid_thw)

        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len, -1)
        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)

        deepstack_feature_lists = []
        for layer_num, blk in enumerate(self.blocks):
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens,
                position_embeddings=position_embeddings,
                **kwargs,
            )
            if layer_num in self.deepstack_visual_indexes:
                deepstack_feature = self.deepstack_merger_list[
                    self.deepstack_visual_indexes.index(layer_num)
                ](hidden_states)
                deepstack_feature_lists.append(deepstack_feature)

        # hidden_states = self.merger(hidden_states)

        return hidden_states, deepstack_feature_lists


class SMBVisionPredictor(PreTrainedModel):
    config: SMBVisionPredictorConfig
    _no_split_modules = ["SMBVisionBlock"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
    _supports_attention_backend = True

    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)

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

        self.blocks = nn.ModuleList(
            [SMBVisionBlock(config) for _ in range(config.depth)]
        )

        self.in_proj = nn.Linear(config.in_hidden_size, config.hidden_size)
        self.out_proj = nn.Linear(config.hidden_size, config.in_hidden_size)
        self.mask_token = nn.Parameter(torch.zeros(config.hidden_size))
        self.gradient_checkpointing = False

    def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
        merge_size = 1

        max_hw = int(grid_thw[:, 1:].max().item())
        freq_table = self.rotary_pos_emb(max_hw)  # (max_hw, dim // 2)
        device = freq_table.device

        total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
        pos_ids = torch.empty((total_tokens, 2), dtype=torch.long, device=device)

        offset = 0
        for num_frames, height, width in grid_thw:
            merged_h, merged_w = height // merge_size, width // merge_size

            block_rows = torch.arange(merged_h, device=device)  # block row indices
            block_cols = torch.arange(merged_w, device=device)  # block col indices
            intra_row = torch.arange(
                merge_size, device=device
            )  # intra-block row offsets
            intra_col = torch.arange(
                merge_size, device=device
            )  # intra-block col offsets

            # Compute full-resolution positions
            row_idx = (
                block_rows[:, None, None, None] * merge_size
                + intra_row[None, None, :, None]
            )
            col_idx = (
                block_cols[None, :, None, None] * merge_size
                + intra_col[None, None, None, :]
            )

            row_idx = row_idx.expand(
                merged_h, merged_w, merge_size, merge_size
            ).reshape(-1)
            col_idx = col_idx.expand(
                merged_h, merged_w, merge_size, merge_size
            ).reshape(-1)

            coords = torch.stack((row_idx, col_idx), dim=-1)

            if num_frames > 1:
                coords = coords.repeat(num_frames, 1)

            num_tokens = coords.shape[0]
            pos_ids[offset : offset + num_tokens] = coords
            offset += num_tokens

        embeddings = freq_table[pos_ids]  # lookup rotary embeddings
        embeddings = embeddings.flatten(1)
        return embeddings

    def forward(
        self,
        hidden_states: torch.Tensor,
        grid_thw: torch.Tensor,
        target_mask: torch.Tensor,
        **kwargs,
    ) -> torch.Tensor:
        # mask out the hidden states
        hidden_states = self.in_proj(hidden_states)
        hidden_states[target_mask] = self.mask_token

        # apply position embeddings
        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        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):
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        # hidden_states = self.merger(hidden_states)
        hidden_states = self.out_proj(hidden_states)

        return hidden_states


@dataclass
class SMBVisionModelOutput(ModelOutput):
    loss: Optional[torch.FloatTensor] = None
    mim_loss: Optional[torch.FloatTensor] = None
    jepa_loss: Optional[torch.FloatTensor] = None
    hidden_states: Optional[torch.FloatTensor] = None
    enc_hidden_states: Optional[torch.FloatTensor] = None
    predicted_hidden_states: Optional[torch.FloatTensor] = None


class SMBVisionPretrainedModel(PreTrainedModel):
    config: SMBVisionModelConfig
    base_model_prefix = ""
    supports_gradient_checkpointing = True
    _no_split_modules = ["SMBVisionBlock"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
    _supports_attention_backend = True

    def _init_weights(self, module):
        """Initialize the weights"""

        init_std = self.config.vision_config.initializer_range

        # Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
        # `trunc_normal_cpu` not implemented in `half` issues
        def trunc_normal_f32_(weight, std):
            data_float_32 = weight.data.to(torch.float32)
            data_init = nn.init.trunc_normal_(data_float_32, mean=0.0, std=std)
            weight.data = data_init.to(weight.dtype)

        if isinstance(module, SMBVisionEncoder):
            trunc_normal_f32_(module.pos_embed.weight, std=init_std)
            for i, layer in enumerate(module.blocks, 1):
                std = init_std / (i**0.5)
                trunc_normal_f32_(layer.attn.proj.weight, std=std)
                trunc_normal_f32_(layer.mlp.fc2.weight, std=std)
            std = init_std / (len(module.blocks) + 1) ** 0.5
            trunc_normal_f32_(module.mlp.fc2.weight, std=std)
        elif isinstance(module, SMBVisionPredictor):
            trunc_normal_f32_(module.mask_token, std=init_std)
            trunc_normal_f32_(module.in_proj.weight, std=init_std)
            trunc_normal_f32_(module.out_proj.weight, std=init_std)
            for i, layer in enumerate(module.blocks, 1):
                std = init_std / (i**0.5)
                trunc_normal_f32_(layer.attn.proj.weight, std=std)
                trunc_normal_f32_(layer.mlp.fc2.weight, std=std)
            std = init_std / (len(module.blocks) + 1) ** 0.5
            trunc_normal_f32_(module.mlp.fc2.weight, std=std)
        elif isinstance(module, (nn.Linear, nn.Conv2d, nn.Conv3d)):
            trunc_normal_f32_(module.weight, std=init_std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


class SMBVisionModel(SMBVisionPretrainedModel):
    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)

        self.encoder = SMBVisionEncoder._from_config(config.vision_config)
        self.predictor = SMBVisionPredictor._from_config(config.predictor_config)
        self.to_pixels = nn.Linear(
            config.vision_config.hidden_size,
            config.vision_config.patch_size**2
            * config.vision_config.temporal_patch_size,
        )
        self.masking_ratio = config.masking_ratio
        self.mask_token = nn.Parameter(
            torch.zeros(
                config.vision_config.in_channels
                * config.vision_config.temporal_patch_size
                * config.vision_config.patch_size**2
            )
        )

        self.mim_loss = nn.L1Loss(reduction="mean")
        self.jepa_loss = nn.MSELoss(reduction="mean")

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

    def forward_features(
        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs
    ) -> torch.Tensor:
        return self.encoder(hidden_states, grid_thw, **kwargs)

    def forward(
        self,
        hidden_states: torch.Tensor,
        grid_thw: torch.Tensor,
        context_mask: Optional[torch.Tensor],
        target_mask: Optional[torch.Tensor],
        **kwargs,
    ) -> torch.Tensor:
        # modeling masked image reconstruction
        # prepare mask tokens
        num_masked = int(self.masking_ratio * hidden_states.shape[0])
        masked_indices = torch.randperm(hidden_states.shape[0])[:num_masked]
        # replace masked indices with mask tokens
        inputs_mim = hidden_states.clone()
        inputs_mim[masked_indices] = self.mask_token.to(hidden_states.dtype)
        masked_hidden_states, deepstack_feature_lists = self.encoder(
            inputs_mim, grid_thw, **kwargs
        )
        masked_hidden_states = self.to_pixels(masked_hidden_states)
        # compute mim loss
        mim_loss = self.mim_loss(
            masked_hidden_states[masked_indices], hidden_states[masked_indices]
        )

        # modeling next embedding prediction
        if context_mask is not None and target_mask is not None:
            context_mask = context_mask == 1
            target_mask = target_mask == 1
            # extend context and target masks
            lengths = torch.prod(grid_thw, dim=1)
            extended_context_mask = torch.repeat_interleave(context_mask, lengths)
            extended_target_mask = torch.repeat_interleave(target_mask, lengths)

            enc_hidden_states, deepstack_feature_lists = self.encoder(
                hidden_states[extended_context_mask], grid_thw[context_mask], **kwargs
            )
            pred_hidden_states = self.predictor(
                enc_hidden_states,
                grid_thw[context_mask],
                extended_target_mask,
                **kwargs,
            )
            jepa_loss = self.jepa_loss(
                pred_hidden_states[extended_target_mask],
                enc_hidden_states[extended_target_mask],
            )

            loss = mim_loss + jepa_loss
            return SMBVisionModelOutput(
                loss=loss,
                mim_loss=mim_loss,
                jepa_loss=jepa_loss,
                hidden_states=hidden_states,
                enc_hidden_states=enc_hidden_states,
                predicted_hidden_states=pred_hidden_states,
            )
        else:
            return SMBVisionModelOutput(
                loss=mim_loss,
                mim_loss=mim_loss,
                jepa_loss=None,
                hidden_states=hidden_states,
                predicted_hidden_states=None,
            )


__all__ = ["SMBVisionEncoder", "SMBVisionPredictor", "SMBVisionModel"]