qwen3_moe_vit.py

# Copyright 2025 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch Qwen3_ViT model."""

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

from transformers.activations import ACT2FN

from transformers.utils import logging

from typing import Any, Callable, Optional, Union
from functools import partial

from transformers.configuration_utils import PretrainedConfig
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel

logger = logging.get_logger(__name__)


class Qwen3VLMoeVisionConfig(PretrainedConfig):
    model_type = "qwen3_moe_vit"
    
    def __init__(
        self,
        depth=27,
        hidden_size=1152,
        hidden_act="gelu_pytorch_tanh",
        intermediate_size=4304,
        num_heads=16,
        in_channels=3,
        patch_size=16,
        spatial_merge_size=2,
        temporal_patch_size=2,
        out_hidden_size=3584,
        num_position_embeddings=2304,
        deepstack_visual_indexes=[8, 16, 24],
        initializer_range=0.02,
        **kwargs,
    ):
        super().__init__(**kwargs)
        
        self.depth = depth
        self.hidden_size = hidden_size
        self.hidden_act = hidden_act
        self.intermediate_size = intermediate_size
        self.num_heads = num_heads
        self.in_channels = in_channels
        self.patch_size = patch_size
        self.spatial_merge_size = spatial_merge_size
        self.temporal_patch_size = temporal_patch_size
        self.out_hidden_size = out_hidden_size
        self.num_position_embeddings = num_position_embeddings
        self.initializer_range = initializer_range
        self.deepstack_visual_indexes = deepstack_visual_indexes
    
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        cls._set_token_in_kwargs(kwargs)
        
        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
        
        if 'vision_config' in config_dict:
            config_dict = config_dict['vision_config']
        
        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
            logger.warning(
                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
            )
        
        return cls.from_dict(config_dict, **kwargs)


class GradientCheckpointingLayer(nn.Module):
    """Base class for layers with gradient checkpointing.

    This class enables gradient checkpointing functionality for a layer. By default, gradient checkpointing is disabled
    (`gradient_checkpointing = False`). When `model.set_gradient_checkpointing()` is called, gradient checkpointing is
    enabled by setting `gradient_checkpointing = True` and assigning a checkpointing function to `_gradient_checkpointing_func`.

    Important:

        When using gradient checkpointing with `use_reentrant=True`, inputs that require gradients (e.g. hidden states)
        must be passed as positional arguments (`*args`) rather than keyword arguments to properly propagate gradients.

        Example:

            ```python
            >>> # Correct - hidden_states passed as positional arg
            >>> out = self.layer(hidden_states, attention_mask=attention_mask)

            >>> # Incorrect - hidden_states passed as keyword arg
            >>> out = self.layer(hidden_states=hidden_states, attention_mask=attention_mask)
            ```
    """
    
    gradient_checkpointing = False
    
    def __call__(self, *args, **kwargs):
        if self.gradient_checkpointing and self.training:
            do_warn = False
            layer_name = self.__class__.__name__
            message = f"Caching is incompatible with gradient checkpointing in {layer_name}. Setting"
            
            if "use_cache" in kwargs and kwargs["use_cache"]:
                kwargs["use_cache"] = False
                message += " `use_cache=False`,"
                do_warn = True
            
            # different names for the same thing in different layers
            # TODO cyril: this one without `S` can be removed after deprection cycle
            if "past_key_value" in kwargs and kwargs["past_key_value"] is not None:
                kwargs["past_key_value"] = None
                message += " `past_key_value=None`,"
                do_warn = True
            
            if "past_key_values" in kwargs and kwargs["past_key_values"] is not None:
                kwargs["past_key_values"] = None
                message += " `past_key_values=None`,"
                do_warn = True
            
            if "layer_past" in kwargs and kwargs["layer_past"] is not None:
                kwargs["layer_past"] = None
                message += " `layer_past=None`,"
                do_warn = True
            
            # warn if anything was changed
            if do_warn:
                message = message.rstrip(",") + "."
                logger.warning_once(message)
            
            return self._gradient_checkpointing_func(partial(super().__call__, **kwargs), *args)
        return super().__call__(*args, **kwargs)


class Qwen3VLMoeVisionMLP(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 Qwen3VLMoeVisionPatchEmbed(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]
        self.proj = nn.Conv3d(self.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.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 Qwen3VLMoeVisionRotaryEmbedding(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 Qwen3VLMoeVisionPatchMerger(nn.Module):
    def __init__(self, config: Qwen3VLMoeVisionConfig, 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.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 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 Qwen3VLMoeVisionAttention(nn.Module):
    def __init__(self, config: Qwen3VLMoeVisionConfig) -> 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
        self.config._attn_implementation = "flash_attention_2"
    
    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        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().item()
            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 Qwen3VLMoeVisionBlock(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 = Qwen3VLMoeVisionAttention(config=config)
        self.mlp = Qwen3VLMoeVisionMLP(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 Qwen3MoeVisionTransformer(PreTrainedModel):
    config: Qwen3VLMoeVisionConfig
    _no_split_modules = ["Qwen3VLMoeVisionBlock"]
    
    def __init__(self, config, use_deepstack=False, *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.use_deepstack = use_deepstack
        
        self.patch_embed = Qwen3VLMoeVisionPatchEmbed(
            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 = Qwen3VLMoeVisionRotaryEmbedding(head_dim // 2)
        
        self.blocks = nn.ModuleList([Qwen3VLMoeVisionBlock(config) for _ in range(config.depth)])
        self.merger = Qwen3VLMoeVisionPatchMerger(
            config=config,
            use_postshuffle_norm=False,
        )
        
        self.deepstack_visual_indexes = config.deepstack_visual_indexes
        self.deepstack_merger_list = nn.ModuleList(
            [
                Qwen3VLMoeVisionPatchMerger(
                    config=config,
                    use_postshuffle_norm=True,
                )
                for _ in range(len(config.deepstack_visual_indexes))
            ]
        )
        
        self.gradient_checkpointing = False
        self.image_emb_dim = config.out_hidden_size
    
    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
                )
                if self.use_deepstack:
                    deepstack_feature_lists.append(deepstack_feature)
        
        hidden_states = self.merger(hidden_states)
        if self.use_deepstack:
            return hidden_states, deepstack_feature_lists
        return hidden_states