#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/qwen3_vl_moe/modular_qwen3_vl_moe.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_qwen3_vl_moe.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# 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.

import itertools
from collections.abc import Callable
from dataclasses import dataclass

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernelized_func
from ...masking_utils import create_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPast,
    BaseModelOutputWithPooling,
    MoeCausalLMOutputWithPast,
    MoeModelOutputWithPast,
)
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, slice_logits_to_keep
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check
from ...utils.generic import (
    accepts_precomputed_kwargs,
    get_max_seqlen,
    is_flash_attention_requested,
    merge_with_config_defaults,
)
from ...utils.output_capturing import OutputRecorder, capture_outputs
from ...vision_utils import (
    get_vision_attention_seqlens,
    get_vision_interpolation_indices_and_weights,
    get_vision_position_ids,
)
from ..auto.modeling_auto import AutoModel
from .configuration_qwen3_vl_moe import Qwen3VLMoeConfig, Qwen3VLMoeTextConfig, Qwen3VLMoeVisionConfig


@use_kernel_forward_from_hub("RMSNorm")
class Qwen3VLMoeTextRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        Qwen3VLMoeTextRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


@use_experts_implementation
class Qwen3VLMoeTextExperts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.moe_intermediate_size
        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(
        self,
        hidden_states: torch.Tensor,
        top_k_index: torch.Tensor,
        top_k_weights: torch.Tensor,
    ) -> torch.Tensor:
        final_hidden_states = torch.zeros_like(hidden_states)
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts + 1)
            expert_mask = expert_mask.permute(2, 1, 0)
            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()

        for expert_idx in expert_hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
            current_state = hidden_states[token_idx]
            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
            current_hidden_states = self.act_fn(gate) * up
            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))

        return final_hidden_states


class Qwen3VLMoeTextTopKRouter(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_experts
        self.hidden_dim = config.hidden_size
        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))

    def forward(self, hidden_states):
        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
        router_logits = F.linear(hidden_states, self.weight)  # (seq_len, num_experts)
        router_probs = torch.nn.functional.softmax(router_logits, dtype=torch.float, dim=-1)
        router_top_value, router_indices = torch.topk(router_probs, self.top_k, dim=-1)  # (seq_len, top_k)
        router_top_value /= router_top_value.sum(dim=-1, keepdim=True)
        router_top_value = router_top_value.to(router_logits.dtype)
        router_scores = router_top_value
        return router_logits, router_scores, router_indices


class Qwen3VLMoeTextSparseMoeBlock(nn.Module):
    def __init__(self, config: Qwen3VLMoeTextConfig):
        super().__init__()
        self.experts = Qwen3VLMoeTextExperts(config)
        self.gate = Qwen3VLMoeTextTopKRouter(config)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states_reshaped = hidden_states.view(-1, hidden_dim)
        _, routing_weights, selected_experts = self.gate(hidden_states_reshaped)
        final_hidden_states = self.experts(hidden_states_reshaped, selected_experts, routing_weights)
        return final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)


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: torch.Tensor | None,
    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:
        attn_weights = attn_weights + attention_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


@use_kernel_forward_from_hub("rotary_pos_emb")
def apply_rotary_pos_emb(q, k, cos, sin, 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.
        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


@use_kernelized_func(apply_rotary_pos_emb)
class Qwen3VLMoeTextAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Qwen3VLMoeTextConfig, layer_idx: int):
        super().__init__()
        self.layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.q_norm = Qwen3VLMoeTextRMSNorm(
            self.head_dim, eps=config.rms_norm_eps
        )  # unlike olmo, only on the head dim!
        self.k_norm = Qwen3VLMoeTextRMSNorm(
            self.head_dim, eps=config.rms_norm_eps
        )  # thus post q_norm does not need reshape

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        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,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Qwen3VLMoeTextMLP(nn.Module):
    def __init__(self, config, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else 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


class Qwen3VLMoeTextDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Qwen3VLMoeTextConfig, layer_idx: int):
        super().__init__()
        self.self_attn = Qwen3VLMoeTextAttention(config, layer_idx)
        if (layer_idx not in config.mlp_only_layers) and (
            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
        ):
            self.mlp = Qwen3VLMoeTextSparseMoeBlock(config)
        else:
            self.mlp = Qwen3VLMoeTextMLP(config, intermediate_size=config.intermediate_size)
        self.input_layernorm = Qwen3VLMoeTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen3VLMoeTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.hidden_size = config.hidden_size

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states


@auto_docstring
class Qwen3VLMoePreTrainedModel(PreTrainedModel):
    config: Qwen3VLMoeConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen3VLMoeTextDecoderLayer", "Qwen3VLMoeVisionBlock"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "router_logits": OutputRecorder(Qwen3VLMoeTextTopKRouter, index=0),
        "hidden_states": Qwen3VLMoeTextDecoderLayer,
        "attentions": Qwen3VLMoeTextAttention,
    }
    input_modalities = ("text", "image", "video")

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights."""
        super()._init_weights(module)
        if hasattr(self.config, "initializer_range"):
            std = self.config.initializer_range
        else:
            std = getattr(self.config.get_text_config(), "initializer_range", 0.02)
        if isinstance(module, Qwen3VLMoeTextExperts):
            init.normal_(module.gate_up_proj, mean=0.0, std=std)
            init.normal_(module.down_proj, mean=0.0, std=std)
        elif isinstance(module, Qwen3VLMoeTextTopKRouter):
            init.normal_(module.weight, mean=0.0, std=std)


class Qwen3VLMoeVisionRotaryEmbedding(nn.Module):
    """
    Simple axial 2D rope with same freqs used for H and W grids. The freqs are
    pre-computed using `head-dim//4` which is later used to concat H and W positions.
    The final angles rotate over the whole head dim, no partial rotation involved.
    """

    def __init__(self, config: Qwen3VLMoeVisionConfig):
        super().__init__()
        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_axial_rope_parameters
        if self.rope_type != "axial":
            raise ValueError(f"{self.__class__.__name__} supports only axial rope, but requested {self.rope_type}")
        inv_freq, self.attention_scaling = rope_init_fn(self.config)

        self.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)

    @staticmethod
    def compute_axial_rope_parameters(
        config: Qwen3VLMoeVisionConfig, device=None, **kwargs
    ) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        spatial_dim = dim // 2

        attention_factor = 1.0  # Unused in this type of RoPE
        inv_freq = 1.0 / (base ** (torch.arange(0, spatial_dim, 2, dtype=torch.float) / spatial_dim))
        return inv_freq.to(device), attention_factor

    @torch.no_grad()
    def forward(self, x, position_ids):
        # position_ids: (2, N) — row 0 = h coords, row 1 = w coords
        freqs = position_ids[..., None].float() * self.inv_freq.to(device=x.device, dtype=torch.float)
        cos = freqs.cos() * self.attention_scaling
        sin = freqs.sin() * self.attention_scaling

        cos = self.recomposition_frequencies(cos)
        sin = self.recomposition_frequencies(sin)
        return cos, sin

    def recomposition_frequencies(self, freq):
        """
        Recompose the frequencies into the final spatial layout used per each grid.
        """
        freq_h, freq_w = freq[:, 0], freq[:, 1]
        freq_hw = torch.cat([freq_h, freq_w], dim=-1)
        return torch.cat([freq_hw, freq_hw], 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


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

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        max_seqlen: int | None = 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 = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        if is_flash_attention_requested(self.config):
            # Flash Attention: Use cu_seqlens for variable length attention
            max_seqlen = get_max_seqlen(cu_seqlens, self.config, kwargs={"max_seqlen": max_seqlen})
            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 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 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)

    @auto_docstring
    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs,
    ) -> torch.Tensor:
        r"""
        cu_seqlens (`torch.Tensor`):
            Cumulative sequence lengths used for packed variable-length attention in Flash Attention kernels.
        """
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states),
            cu_seqlens=cu_seqlens,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


@auto_docstring
@dataclass
class BaseModelOutputWithDeepstackFeatures(BaseModelOutputWithPooling):
    r"""
    deepstack_features (`List[torch.Tensor]`, *optional*):
        List of hidden-states (feature maps) from deepstack layers.
    """

    deepstack_features: list[torch.Tensor] | None = None


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 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


class Qwen3VLMoeVisionModel(Qwen3VLMoePreTrainedModel):
    config: Qwen3VLMoeVisionConfig
    input_modalities = ("image", "video")
    _can_record_outputs = {
        "router_logits": OutputRecorder(Qwen3VLMoeTextTopKRouter, layer_name="mlp.gate", index=0),
        "hidden_states": Qwen3VLMoeVisionBlock,
        "attentions": Qwen3VLMoeVisionAttention,
    }

    def __init__(self, config: Qwen3VLMoeVisionConfig, *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 = Qwen3VLMoeVisionPatchEmbed(
            config=config,
        )

        self.pos_embed = nn.Embedding(config.num_position_embeddings, config.hidden_size)
        # How the (square) learned position grid is resampled to each image's grid.
        self.num_grid_per_side = int(config.num_position_embeddings**0.5)
        self.interpolation_align_corners = True
        self.interpolation_mode = "bilinear"

        self.rotary_pos_emb = Qwen3VLMoeVisionRotaryEmbedding(config)

        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.post_init()

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
        """
        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.
        """
        interp_indices, interp_weights = get_vision_interpolation_indices_and_weights(
            grid_thw,
            num_grid_per_side=self.num_grid_per_side,
            mode=self.interpolation_mode,
            align_corners=self.interpolation_align_corners,
            spatial_merge_size=self.config.spatial_merge_size,
            kwargs=kwargs,
        )
        position_ids = get_vision_position_ids(grid_thw, self.spatial_merge_size, kwargs=kwargs)
        cu_seqlens, max_seqlen = get_vision_attention_seqlens(grid_thw, self.config, kwargs=kwargs)

        hidden_states = self.patch_embed(hidden_states)
        pos_embeds = (self.pos_embed(interp_indices) * interp_weights[:, :, None]).sum(1)
        hidden_states = hidden_states + pos_embeds.to(hidden_states.dtype)
        position_embeddings = self.rotary_pos_emb(hidden_states, position_ids)

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

        deepstack_feature_lists = []
        for layer_num, blk in enumerate(self.blocks):
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens,
                max_seqlen=max_seqlen,
                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)

        merged_hidden_states = self.merger(hidden_states)

        return BaseModelOutputWithDeepstackFeatures(
            last_hidden_state=hidden_states,
            pooler_output=merged_hidden_states,
            deepstack_features=deepstack_feature_lists,
        )


class Qwen3VLMoeTextRotaryEmbedding(nn.Module):
    def __init__(self, config: Qwen3VLMoeTextConfig):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_default_rope_parameters
        if self.rope_type != "default":
            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = rope_init_fn(self.config)

        self.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)
        self.mrope_section = config.rope_parameters.get("mrope_section", [24, 20, 20])

    @staticmethod
    def compute_default_rope_parameters(
        config: Qwen3VLMoeTextConfig, device=None, **kwargs
    ) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

        attention_factor = 1.0  # Unused in this type of RoPE
        # Compute the inverse frequencies
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        return inv_freq, attention_factor

    @torch.no_grad()
    @dynamic_rope_update
    def forward(self, x, position_ids):
        # In contrast to other models, Qwen3VLMoeText has different position ids for the grids
        # So position_ids is broadcast to (3, bs, positions) and freqs has shape (3, bs, positions, dim // 2)
        position_ids = position_ids.expand(3, -1, -1)
        freqs = position_ids[..., None].float() * self.inv_freq.to(device=x.device, dtype=torch.float)
        cos = freqs.cos() * self.attention_scaling
        sin = freqs.sin() * self.attention_scaling

        sin = self.recomposition_frequencies(sin)
        cos = self.recomposition_frequencies(cos)
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

    def recomposition_frequencies(self, freq):
        """
        Recompose the frequencies into the final spatial layout used per each grid.
        """
        freqs_thw = freq[0]  # just overwrite the first dimension T
        for dim, offset in enumerate((1, 2), start=1):  # H, W
            length = self.mrope_section[dim] * 3
            idx = slice(offset, length, 3)
            freqs_thw[..., idx] = freq[dim, ..., idx]
        return torch.cat((freqs_thw, freqs_thw), dim=-1)


@auto_docstring(
    custom_intro=(
        "Text part of Qwen3VLMoe, "
        "not a pure text-only model, as DeepStack integrates visual features into the early hidden states."
    )
)
class Qwen3VLMoeTextModel(Qwen3VLMoePreTrainedModel):
    config: Qwen3VLMoeTextConfig
    input_modalities = ("text",)

    def __init__(self, config: Qwen3VLMoeTextConfig):
        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(
            [Qwen3VLMoeTextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Qwen3VLMoeTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Qwen3VLMoeTextRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        use_cache: bool | None = None,
        # args for deepstack
        visual_pos_masks: torch.Tensor | None = None,
        deepstack_visual_embeds: list[torch.Tensor] | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple | MoeModelOutputWithPast:
        r"""
        visual_pos_masks (`torch.Tensor` of shape `(batch_size, seqlen)`, *optional*):
            The mask of the visual positions.
        deepstack_visual_embeds (`list[torch.Tensor]`, *optional*):
            The deepstack visual embeddings. The shape is (num_layers, visual_seqlen, embed_dim).
            The feature is extracted from the different visual encoder layers, and fed to the decoder
            hidden states. It's from the paper DeepStack(https://arxiv.org/abs/2406.04334).
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        # 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)

        # the hard coded `4` is for text, temporal, height and width.
        if position_ids is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
            position_ids = position_ids.view(1, 1, -1).expand(4, inputs_embeds.shape[0], -1)
        elif position_ids.ndim == 2:
            position_ids = position_ids[None, ...].expand(4, position_ids.shape[0], -1)

        if position_ids.ndim == 3 and position_ids.shape[0] == 4:
            text_position_ids = position_ids[0]
            position_ids = position_ids[1:]
        else:
            text_position_ids = None

        attention_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=text_position_ids,
        )

        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
        for layer_idx, decoder_layer in enumerate(self.layers):
            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_ids=text_position_ids,
                past_key_values=past_key_values,
                position_embeddings=position_embeddings,
                **kwargs,
            )
            hidden_states = layer_outputs

            # add visual features to the hidden states of first several layers
            if deepstack_visual_embeds is not None and layer_idx in range(len(deepstack_visual_embeds)):
                hidden_states = self._deepstack_process(
                    hidden_states,
                    visual_pos_masks,
                    deepstack_visual_embeds[layer_idx],
                )

        hidden_states = self.norm(hidden_states)

        return MoeModelOutputWithPast(  # only diff with Qwen3VLTextModel
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )

    def _deepstack_process(
        self, hidden_states: torch.Tensor, visual_pos_masks: torch.Tensor, visual_embeds: torch.Tensor
    ):
        visual_pos_masks = visual_pos_masks.to(hidden_states.device)
        visual_embeds = visual_embeds.to(hidden_states.device, hidden_states.dtype)
        hidden_states = hidden_states.clone()
        local_this = hidden_states[visual_pos_masks, :] + visual_embeds
        hidden_states[visual_pos_masks, :] = local_this
        return hidden_states


@auto_docstring
class Qwen3VLMoeModel(Qwen3VLMoePreTrainedModel):
    base_model_prefix = "model"
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False

    def __init__(self, config: Qwen3VLMoeConfig):
        super().__init__(config)
        self.visual = AutoModel.from_config(config.vision_config)
        self.language_model = AutoModel.from_config(config.text_config)
        self.rope_deltas = None  # cache rope_deltas here

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

    def get_vision_position_ids(
        self,
        start_position: int,
        grid_thw: list[int, int, int] | torch.Tensor,
        temp_merge_size: int = 1,
        spatial_merge_size: int = 1,
        time_interval: float | torch.Tensor = 1.0,
        device: str | torch.device | None = None,
    ):
        """
        Compute 3D positional indices for vision tokens derived from a single image or video input.

        The positions are generated from the input grid defined by temporal (T), height (H), and
        width (W) dimensions. Temporal and spatial dimensions can be downscaled according to the
        merge sizes used in the vision backbone. The resulting positions are offset by `start_position`.

        Args:
            start_position (`int`):
                Offset added to all computed positional indices.
            grid_thw (`Sequence[int]` or `torch.Tensor` of shape `(3,)`):
                The (T, H, W) grid representing the feature layout of the current image or video after patch embedding.
            temp_merge_size (`int`, *optional*):
                Factor by which the temporal dimension is reduced in the backbone. The temporal grid size is divided
                by this value. Defaults to 1.
            spatial_merge_size (`int`, *optional*):
                Factor by which the spatial dimensions (H and W) are reduced in the backbone. Both H and W are divided
                by this value. Defaults to 1.
            time_interval (`float` or scalar `torch.Tensor`, *optional*, defaults to 1.0):
                Spacing factor applied before quantizing temporal position indices.
            device (`str` or `torch.device`, *optional*):
                Device on which the resulting tensor is allocated. If `None`, uses the current default device.

        Returns:
            torch.LongTensor of shape (3, sequence_length):
                Positional indices for temporal, height, and width dimensions,
                flattened into sequence form and offset by `start_position`.
        """
        llm_grid_t, llm_grid_h, llm_grid_w = (
            grid_thw[0].item() // temp_merge_size,
            grid_thw[1].item() // spatial_merge_size,
            grid_thw[2].item() // spatial_merge_size,
        )

        position_temporal = (torch.arange(llm_grid_t, device=device) * time_interval).long()
        position_height = torch.arange(llm_grid_h, device=device) + start_position
        position_width = torch.arange(llm_grid_w, device=device) + start_position

        T_grid, H_grid, W_grid = torch.meshgrid(position_temporal, position_height, position_width, indexing="ij")
        vision_position_ids = torch.stack([T_grid, H_grid, W_grid], dim=0).reshape(3, -1)
        vision_position_ids[0] += start_position  # must be after time_interval multiply
        return vision_position_ids

    def get_rope_index(
        self,
        input_ids: torch.LongTensor,
        mm_token_type_ids: torch.IntTensor,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Difference from Qwen2VL/Qwen2.5VL's get_rope_index:
        - Since Qwen3.5 use timestamps to separate videos, like <t1> <vision_start> <frame1> <vision_end> <t2> <vision_start> <frame2> <vision_end>, the video_grid_thw should also be split too.

        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.
            mm_token_type_ids (`torch.IntTensor` of shape `(batch_size, sequence_length)`):
                Token type ids matching each modality to a different value in the input sequence, i.e. text (0), image (1), video (2).
            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.
            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)`)
        """

        # Separate video grid thw into multiple grids because timestamps are used to separate videos.
        if video_grid_thw is not None:
            video_grid_thw = torch.repeat_interleave(video_grid_thw, video_grid_thw[:, 0], dim=0)
            video_grid_thw[:, 0] = 1
        spatial_merge_size = self.config.vision_config.spatial_merge_size

        mrope_position_deltas = []
        position_ids = torch.zeros(
            3,
            input_ids.shape[0],
            input_ids.shape[1],
            dtype=input_ids.dtype,
            device=input_ids.device,
        )
        grid_iters = {
            1: iter(image_grid_thw) if image_grid_thw is not None else None,
            2: iter(video_grid_thw) if video_grid_thw is not None else None,
        }

        for batch_idx, current_input_ids in enumerate(input_ids):
            input_token_type = mm_token_type_ids[batch_idx]
            if attention_mask is not None:
                current_input_ids = current_input_ids[attention_mask[batch_idx].bool()]
                input_token_type = input_token_type[attention_mask[batch_idx].bool()]

            input_type_group = []
            for key, group in itertools.groupby(enumerate(input_token_type.tolist()), lambda x: x[1]):
                group = list(group)
                start_index = group[0][0]
                end_index = group[-1][0] + 1
                input_type_group.append((key, start_index, end_index))

            current_pos = 0
            llm_pos_ids_list = []
            for modality_type, start_idx, end_idx in input_type_group:
                # text == 0
                if modality_type == 0:
                    text_len = end_idx - start_idx
                    llm_pos_ids_list.append(
                        torch.arange(text_len, device=input_ids.device).view(1, -1).expand(3, -1) + current_pos
                    )
                    current_pos += text_len
                # image == 1, video == 2
                else:
                    grid_thw = next(grid_iters[modality_type])
                    vision_position_ids = self.get_vision_position_ids(
                        current_pos, grid_thw, 1, spatial_merge_size, device=input_ids.device
                    )
                    llm_pos_ids_list.append(vision_position_ids)
                    current_pos += max(grid_thw[1], grid_thw[2]) // spatial_merge_size
            llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
            if attention_mask is not None:
                position_ids[:, batch_idx, attention_mask[batch_idx].bool()] = llm_positions.to(position_ids.device)
            else:
                position_ids[:, batch_idx] = llm_positions.to(position_ids.device)
            mrope_position_deltas.append(llm_positions.max() + 1 - len(current_input_ids))
        mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
        return position_ids, mrope_position_deltas

    @accepts_precomputed_kwargs(modality="video")
    @can_return_tuple
    @auto_docstring
    def get_video_features(
        self,
        pixel_values_videos: torch.Tensor,
        video_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
        # Same implementation as for images
        return self.get_image_features(pixel_values_videos, video_grid_thw, **kwargs)

    @accepts_precomputed_kwargs(modality="image")
    @can_return_tuple
    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.Tensor,
        image_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
        pixel_values = pixel_values.type(self.visual.dtype)
        vision_output: BaseModelOutputWithDeepstackFeatures = self.visual(
            pixel_values, grid_thw=image_grid_thw, return_dict=True, **kwargs
        )
        image_embeds = vision_output.pooler_output
        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        vision_output.pooler_output = torch.split(image_embeds, split_sizes)
        vision_output.deepstack_features = [
            torch.split(feat, split_sizes) for feat in vision_output.deepstack_features
        ]
        return vision_output

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.Tensor,
        image_features: torch.Tensor | None = None,
        video_features: torch.Tensor | None = 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.full((), 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.full((), 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).to(inputs_embeds.device)
        if image_features is not None:
            torch_compilable_check(
                n_image_tokens * inputs_embeds.shape[-1] == image_features.numel(),
                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).to(inputs_embeds.device)
        if video_features is not None:
            torch_compilable_check(
                n_video_tokens * inputs_embeds.shape[-1] == video_features.numel(),
                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
            )
        return special_image_mask, special_video_mask

    def compute_3d_position_ids(
        self,
        input_ids: torch.Tensor | None,
        inputs_embeds: torch.Tensor | None,
        image_grid_thw: torch.Tensor | None = None,
        video_grid_thw: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        past_key_values: torch.Tensor | None = None,
        mm_token_type_ids: torch.IntTensor | None = None,
    ) -> torch.Tensor | None:
        past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()
        has_multimodal = image_grid_thw is not None or video_grid_thw is not None
        if has_multimodal and mm_token_type_ids is None and input_ids is not None:
            raise ValueError(
                "Multimodal data was passed (via `image_grid_thw` or `video_grid_thw`) but `mm_token_type_ids` is "
                "missing. Please pass `mm_token_type_ids` to the model so that multimodal RoPE (M-RoPE) can be "
                "computed correctly. `mm_token_type_ids` is returned by the processor alongside `input_ids`."
            )
        can_compute_mrope = input_ids is not None and mm_token_type_ids is not None and has_multimodal

        if can_compute_mrope and (self.rope_deltas is None or past_key_values_length == 0):
            position_ids, rope_deltas = self.get_rope_index(
                input_ids,
                image_grid_thw=image_grid_thw,
                video_grid_thw=video_grid_thw,
                attention_mask=attention_mask,
                mm_token_type_ids=mm_token_type_ids,
            )
            self.rope_deltas = rope_deltas
        # Use pre-calculated rope-deltas to infer correct 3D position ids during incremental
        # generation (past_key_values_length > 0) or when only inputs_embeds is provided (no input_ids
        # to recompute from). Skip when input_ids is provided without past_key_values to avoid shape
        # mismatches from stale rope_deltas (e.g., training forward pass after generation).
        elif self.rope_deltas is not None and (past_key_values_length > 0 or input_ids is None):
            batch_size, seq_length, _ = inputs_embeds.shape
            if attention_mask is not None:
                position_ids = attention_mask.long().cumsum(-1) - 1
                position_ids = position_ids.masked_fill(attention_mask == 0, 0)
                position_ids = position_ids.view(1, batch_size, -1).repeat(3, 1, 1).to(inputs_embeds.device)
            else:
                position_ids = torch.arange(past_key_values_length, past_key_values_length + seq_length)
                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1).to(inputs_embeds.device)
            delta = self.rope_deltas.repeat_interleave(batch_size // self.rope_deltas.shape[0], dim=0)
            position_ids = position_ids + delta.to(device=inputs_embeds.device)
        else:
            # Can't build correct 3D positions. Let the model infer it
            position_ids = None
        return position_ids

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        pixel_values: torch.Tensor | None = None,
        pixel_values_videos: torch.Tensor | None = None,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        mm_token_type_ids: torch.IntTensor | None = None,
        mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if (pixel_values is not None or pixel_values_videos is not None) and mm_encoder_outputs is not None:
            raise ValueError(
                "You cannot specify both pixel_values/pixel_values_videos and mm_encoder_outputs at the same time"
            )

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

        image_mask = None
        video_mask = None

        mm_encoder_outputs = mm_encoder_outputs if mm_encoder_outputs is not None else {}
        if mm_encoder_outputs.get("image") is None and pixel_values is not None:
            mm_encoder_outputs["image"]: BaseModelOutputWithDeepstackFeatures = self.get_image_features(
                pixel_values, image_grid_thw, return_dict=True, **kwargs
            )

        if mm_encoder_outputs.get("video") is None and pixel_values_videos is not None:
            mm_encoder_outputs["video"]: BaseModelOutputWithDeepstackFeatures = self.get_video_features(
                pixel_values_videos, video_grid_thw, return_dict=True, **kwargs
            )

        if mm_encoder_outputs.get("image") is not None:
            image_embeds = mm_encoder_outputs["image"].pooler_output
            deepstack_image_embeds = mm_encoder_outputs["image"].deepstack_features
            image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            image_mask, _ = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

        if mm_encoder_outputs.get("video") is not None:
            video_embeds = mm_encoder_outputs["video"].pooler_output
            deepstack_video_embeds = mm_encoder_outputs["video"].deepstack_features
            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)

        visual_pos_masks = None
        deepstack_visual_embeds = None
        if image_mask is not None and video_mask is not None:
            # aggregate visual_pos_masks and deepstack_visual_embeds
            image_mask = image_mask[..., 0]
            video_mask = video_mask[..., 0]
            visual_pos_masks = image_mask | video_mask
            deepstack_visual_embeds = []
            image_mask_joint = image_mask[visual_pos_masks]
            video_mask_joint = video_mask[visual_pos_masks]
            for img_embed, vid_embed in zip(deepstack_image_embeds, deepstack_video_embeds):
                img_embed = torch.cat(img_embed, dim=0)
                vid_embed = torch.cat(vid_embed, dim=0)
                embed_joint = img_embed.new_zeros(visual_pos_masks.sum(), img_embed.shape[-1])

                embed_joint[image_mask_joint, :] = img_embed
                embed_joint[video_mask_joint, :] = vid_embed
                deepstack_visual_embeds.append(embed_joint)
        elif image_mask is not None:
            image_mask = image_mask[..., 0]
            visual_pos_masks = image_mask
            deepstack_visual_embeds = [torch.cat(feat, dim=0) for feat in deepstack_image_embeds]
        elif video_mask is not None:
            video_mask = video_mask[..., 0]
            visual_pos_masks = video_mask
            deepstack_visual_embeds = [torch.cat(feat, dim=0) for feat in deepstack_video_embeds]

        if position_ids is None:
            position_ids = self.compute_3d_position_ids(
                input_ids=input_ids,
                image_grid_thw=image_grid_thw,
                video_grid_thw=video_grid_thw,
                inputs_embeds=inputs_embeds,
                attention_mask=attention_mask,
                past_key_values=past_key_values,
                mm_token_type_ids=mm_token_type_ids,
            )

        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,
            visual_pos_masks=visual_pos_masks,
            deepstack_visual_embeds=deepstack_visual_embeds,
            **kwargs,
        )

        return outputs


def load_balancing_loss_func(
    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
    num_experts: int | None = None,
    top_k=2,
    attention_mask: torch.Tensor | None = None,
) -> torch.Tensor | int:
    r"""
    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.

    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
    experts is too unbalanced.

    Args:
        gate_logits:
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        num_experts:
            Number of experts
        top_k:
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter.
        attention_mask (`torch.Tensor`, *optional*):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.

    Returns:
        The auxiliary loss.
    """
    if gate_logits is None or not isinstance(gate_logits, tuple):
        return 0

    # Accumulate assignment counts and probability sums layer by layer, normalizing at the end,
    # so peak memory stays O(seq_len * num_experts) regardless of the number of layers.
    compute_device = gate_logits[0].device
    # Use histc to avoid bincount's device-to-host sync.
    # GPU requires integers for determinism; CPU and MPS require floats for support and determinism.
    histc_dtype = torch.float32 if compute_device.type in ("cpu", "mps") else torch.int32
    tokens_per_expert_sum = torch.zeros(num_experts, dtype=torch.float32, device=compute_device)
    router_prob_sum = torch.zeros(num_experts, dtype=torch.float32, device=compute_device)
    total_rows = 0.0

    if attention_mask is not None:
        # The same flat mask applies to every layer's [batch_size * sequence_length] rows.
        flat_mask = attention_mask.reshape(-1).to(device=compute_device, dtype=torch.float32)

    for layer_gate in gate_logits:
        routing_weights = torch.nn.functional.softmax(layer_gate.to(compute_device), dim=-1)
        _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
        if attention_mask is None:
            # Count top-k assignments with fixed bounds to avoid a device-to-host sync.
            tokens_per_expert_sum = tokens_per_expert_sum + torch.histc(
                selected_experts.reshape(-1).to(histc_dtype), bins=num_experts, min=0, max=num_experts
            )
            # Sum of routing probabilities per expert
            router_prob_sum = router_prob_sum + routing_weights.float().sum(dim=0)
            total_rows = total_rows + routing_weights.shape[0]
        else:
            # Same reductions, weighted by the attention mask to exclude padding tokens
            tokens_per_expert_sum = tokens_per_expert_sum + torch.zeros(
                num_experts, dtype=torch.float32, device=compute_device
            ).scatter_add_(0, selected_experts.reshape(-1), flat_mask.repeat_interleave(top_k))
            router_prob_sum = router_prob_sum + (routing_weights.float() * flat_mask.unsqueeze(-1)).sum(dim=0)
            total_rows = total_rows + flat_mask.sum()

    tokens_per_expert = tokens_per_expert_sum / total_rows
    router_prob_per_expert = router_prob_sum / total_rows

    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
    return overall_loss * num_experts


class Qwen3VLMoeForConditionalGeneration(Qwen3VLMoePreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False

    def __init__(self, config: Qwen3VLMoeConfig):
        super().__init__(config)
        self.model = Qwen3VLMoeModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)

        self.post_init()

    @auto_docstring
    def get_video_features(
        self,
        pixel_values_videos: torch.Tensor,
        video_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
        r"""
        pixel_values_videos (`torch.Tensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input videos.
        """
        return self.model.get_video_features(pixel_values_videos, video_grid_thw, **kwargs)

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.Tensor,
        image_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
        r"""
        pixel_values (`torch.Tensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input images.
        """
        return self.model.get_image_features(pixel_values, image_grid_thw, **kwargs)

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.LongTensor | None = None,
        pixel_values: torch.Tensor | None = None,
        pixel_values_videos: torch.Tensor | None = None,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        mm_token_type_ids: torch.IntTensor | None = None,
        output_router_logits: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MoeCausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from PIL import Image
        >>> from huggingface_hub.utils import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration

        >>> model = Qwen3VLMoeForConditionalGeneration.from_pretrained("Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto")
        >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-30B-A3B-Instruct")

        >>> messages = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image",
                        "image": "https://huggingface.co/datasets/hf-internal-testing/transformers-synthetic-assets/resolve/main/images/qwen_vl_demo.jpeg",
                    },
                    {"type": "text", "text": "Describe this image in short."},
                ],
            }
        ]

        >>> # Preparation for inference
        >>> inputs = processor.apply_chat_template(
            messages,
            tokenize=True,
            add_generation_prompt=True,
            return_dict=True,
            return_tensors="pt"
        )
        >>> inputs = inputs.to(model.device)

        >>> # Generate
        >>> generated_ids = model.generate(**inputs, max_new_tokens=128)
        >>> generated_ids_trimmed = [
            out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
        ]
        >>> processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "A woman in a plaid shirt sits on a sandy beach at sunset, smiling as she gives a high-five to a yellow Labrador Retriever wearing a harness. The ocean waves roll in the background."
        ```"""

        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.text_config.output_router_logits
        )

        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,
            mm_token_type_ids=mm_token_type_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            mm_encoder_outputs=mm_encoder_outputs,
            output_router_logits=output_router_logits,
            **kwargs,
        )

        hidden_states = outputs[0]

        hidden_states = slice_logits_to_keep(hidden_states, logits_to_keep)
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            loss = self.loss_function(
                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
            )

        aux_loss = None
        if output_router_logits:
            aux_loss = load_balancing_loss_func(
                outputs.router_logits,
                self.config.text_config.num_experts,
                self.config.text_config.num_experts_per_tok,
                attention_mask,
            )
            if labels is not None:
                loss += self.config.text_config.router_aux_loss_coef * aux_loss.to(
                    loss.device
                )  # make sure to reside in the same device

        return MoeCausalLMOutputWithPast(
            loss=loss,
            aux_loss=aux_loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
        )

    def _prepare_position_ids_for_generation(self, inputs_tensor, model_kwargs):
        # Overwritten -- requires 3D position ids

        text_positions = super()._prepare_position_ids_for_generation(inputs_tensor, model_kwargs)

        # Early exit in case we are continuing generation from past kv. The stored `rope_deltas` come from an earlier
        # call, so if the uncached tokens hold new images or videos their 3D positions have to be computed below.
        past_length = 0
        if (cache := model_kwargs.get("past_key_values")) is not None:
            past_length = cache.get_seq_length()
        mm_token_type_ids = model_kwargs.get("mm_token_type_ids")
        has_new_vision_tokens = mm_token_type_ids is not None and bool(mm_token_type_ids[:, past_length:].any())
        if past_length != 0 and self.model.rope_deltas is not None and not has_new_vision_tokens:
            position_ids = text_positions[None, ...] + self.model.rope_deltas
            return position_ids

        # Otherwise compute 3d position ids for vision tokens and concat with text position ids
        if "input_ids" in model_kwargs and model_kwargs["input_ids"].shape[1] > 0:
            inputs_tensor = model_kwargs["input_ids"]

        is_input_ids = len(inputs_tensor.shape) == 2 and inputs_tensor.dtype in [torch.int, torch.long]
        if (
            is_input_ids
            and model_kwargs.get("mm_token_type_ids") is not None
            and (model_kwargs.get("image_grid_thw") is not None or model_kwargs.get("video_grid_thw") is not None)
        ):
            model_kwargs = {k: v for k, v in model_kwargs.items() if k != "input_ids"}
            vision_positions, rope_deltas = self.model.get_rope_index(inputs_tensor, **model_kwargs)
            self.model.rope_deltas = rope_deltas
        else:
            vision_positions = text_positions.unsqueeze(0).expand(3, -1, -1)
            self.model.rope_deltas = torch.zeros(
                inputs_tensor.shape[0], 1, dtype=torch.long, device=inputs_tensor.device
            )

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

        return position_ids


__all__ = [
    "Qwen3VLMoeVisionModel",
    "Qwen3VLMoeForConditionalGeneration",
    "Qwen3VLMoeModel",
    "Qwen3VLMoePreTrainedModel",
    "Qwen3VLMoeTextModel",
]
