modeling_gemma4.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/gemma4/modular_gemma4.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_gemma4.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 the HuggingFace 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 math
from collections.abc import Callable
from dataclasses import dataclass
from functools import cached_property
from typing import Optional

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

try:
    from torch.nn.attention.flex_attention import (
        flex_attention as _torch_flex_attention,
        create_block_mask as _torch_create_block_mask,
    )
    _FLEX_ATTENTION_AVAILABLE = True
except ImportError:
    _FLEX_ATTENTION_AVAILABLE = False

_FLEX_ATTENTION_MODE = "hybrid"


def set_flex_attention_mode(mode: str):
    global _FLEX_ATTENTION_MODE
    assert mode in ("eager", "flex", "hybrid", "verify"), f"Invalid mode: {mode!r}. Use 'eager', 'flex', 'hybrid', or 'verify'."
    _FLEX_ATTENTION_MODE = mode


def get_flex_attention_mode() -> str:
    return _FLEX_ATTENTION_MODE

from transformers import initialization as init
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.configuration_utils import PreTrainedConfig
from transformers.generation import GenerationMixin
from transformers.integrations import use_experts_implementation  # MoE fused impl KEPT
from transformers.masking_utils import (
    create_bidirectional_mask,
    create_causal_mask,
    create_masks_for_generate,
    create_sliding_window_causal_mask,
)
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import ModelOutput, TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check
from transformers.utils.output_capturing import OutputRecorder
from transformers.models.auto.modeling_auto import AutoModel
from transformers.models.gemma4.configuration_gemma4 import Gemma4AudioConfig, Gemma4Config, Gemma4TextConfig, Gemma4VisionConfig


# ============================================================================
# Local shims — minimal-diff patch over modeling_gemma4_5.5.0.py.
#
# The rest of this file is byte-for-byte identical to the 5.5.0 original,
# with ONLY the following fused/Hub-dispatch symbols neutralized:
#
#   * `use_kernelized_func`  -> no-op class decorator
#     (prevents Hub RoPE kernel from replacing apply_rotary_pos_emb inside
#     Gemma4TextAttention / Gemma4VisionAttention)
#   * `ALL_ATTENTION_FUNCTIONS`  -> import kept (Vision path still uses it),
#     but the single reference inside Gemma4TextAttention.forward is
#     patched to call eager_attention_forward directly.
#   * `maybe_autocast`  -> aliased to torch.autocast (equivalent with
#     enabled=False, but removes the maybe_* fast-path indirection)
#   * `merge_with_config_defaults`  -> no-op function decorator
#     (stops config-default auto-merging on Gemma4TextModel.forward)
#   * `capture_outputs`  -> no-op function decorator
#     (stops forward-hook-based auto collection of hidden_states /
#     router_logits; the distillation wrapper in hybrid_model.py puts a
#     non-tensor payload at position [1] of self_attn's return value and
#     must not be overwritten by hook captures)
#
# Everything else — vision / audio / multimodal classes, KV sharing,
# SWA + sliding-window mask, per-layer-type RoPE (default + proportional),
# per-layer inputs, final logit softcapping, MoE — is UNTOUCHED.
# ============================================================================

# `torch.autocast(enabled=False)` is the exact behavior we want.
maybe_autocast = torch.autocast


def merge_with_config_defaults(fn):
    """No-op replacement. Keeps forward signatures honest; no config-default
    auto-merging of kwargs."""
    return fn


def capture_outputs(fn):
    """No-op replacement. The original installs forward hooks on every
    module listed in `_can_record_outputs` and overwrites output fields
    from captured tensors — which collides with the hybrid_model.py
    AttentionWrapper that stores a (student_hs, teacher_hs) tuple at
    position [1] of the self_attn return. Disabling this ensures the
    explicit `all_self_attns` / `all_hidden_states` collection inside
    Gemma4TextModel.forward is the sole source of truth."""
    return fn


def use_kernelized_func(_fn):
    """No-op class decorator. Prevents the Hub-registered
    `apply_rotary_pos_emb` kernel from being dynamically bound inside
    Gemma4TextAttention / Gemma4VisionAttention — teacher RoPE now runs
    the plain Python implementation on every call."""
    def _decorator(cls):
        return cls
    return _decorator


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Gemma4 outputs, with hidden states and attentions.
    """
)
class Gemma4ModelOutputWithPast(BaseModelOutputWithPast):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    audio_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        audio_hidden_states of the model produced by the audio encoder and after projecting the last hidden state.
    """

    image_hidden_states: torch.FloatTensor | None = None

    audio_hidden_states: torch.FloatTensor | None = None


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Gemma4 causal language model (or autoregressive) outputs.
    """
)
class Gemma4CausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.text_config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder after projecting last hidden state.
    audio_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        audio_hidden_states of the model produced by the audio encoder and after projecting the last hidden state.
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: torch.FloatTensor | None = None

    audio_hidden_states: torch.FloatTensor | None = None


@dataclass
@auto_docstring
class Gemma4AudioModelOutput(BaseModelOutputWithPooling):
    r"""
    attention_mask (`torch.BoolTensor`, *optional*):
        A torch.BoolTensor of shape `(batch_size, num_frames)`. True for valid positions, False for padding.
    """

    attention_mask: torch.BoolTensor | None = None


class Gemma4ClippableLinear(nn.Module):
    def __init__(
        self,
        config: Gemma4VisionConfig | Gemma4AudioConfig,
        in_features: int,
        out_features: int,
    ) -> None:
        super().__init__()
        self.use_clipped_linears = config.use_clipped_linears
        self.linear = nn.Linear(in_features, out_features, bias=False)

        if self.use_clipped_linears:
            self.register_buffer("input_min", torch.tensor(-float("inf")))
            self.register_buffer("input_max", torch.tensor(float("inf")))
            self.register_buffer("output_min", torch.tensor(-float("inf")))
            self.register_buffer("output_max", torch.tensor(float("inf")))

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        if self.use_clipped_linears:
            hidden_states = torch.clamp(hidden_states, self.input_min, self.input_max)

        hidden_states = self.linear(hidden_states)

        if self.use_clipped_linears:
            hidden_states = torch.clamp(hidden_states, self.output_min, self.output_max)

        return hidden_states


class Gemma4RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6, with_scale: bool = True):
        super().__init__()
        self.eps = eps
        self.with_scale = with_scale

        if self.with_scale:
            self.weight = nn.Parameter(torch.ones(dim), requires_grad=True)

    def _norm(self, hidden_states: torch.Tensor):
        mean_squared = hidden_states.pow(2).mean(-1, keepdim=True) + self.eps
        # Use torch.pow() (over torch.sqrt() or torch.rsqrt()) to addess compiler differences between Torch and JAX
        return hidden_states * torch.pow(mean_squared, -0.5)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        normed_output = self._norm(hidden_states.float())
        if self.with_scale:
            normed_output = normed_output * self.weight.float()
        return normed_output.type_as(hidden_states)


class Gemma4AudioRelPositionalEncoding(nn.Module):
    """Sinusoidal relative positional encoding for the audio encoder.

    Produces position embeddings of shape [1, 2*context_size - 1, hidden_size] with
    concatenated [sin..., costransformers.] layout matching the original Gemma4 convention.
    """

    inv_timescales: torch.Tensor

    def __init__(self, config: Gemma4AudioConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.context_size = (
            config.attention_chunk_size + config.attention_context_left - 1 + config.attention_context_right
        )
        min_timescale = 1.0
        max_timescale = 10000.0
        num_timescales = self.hidden_size // 2
        log_timescale_increment = math.log(max_timescale / min_timescale) / max(num_timescales - 1, 1)
        inv_timescales = min_timescale * torch.exp(torch.arange(num_timescales) * -log_timescale_increment)
        self.register_buffer("inv_timescales", inv_timescales.unsqueeze(0).unsqueeze(0), persistent=False)

    @torch.no_grad()
    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        position_ids = torch.arange(12, -1, -1, device=hidden_states.device)
        position_ids = position_ids[..., None]
        scaled_time = position_ids * self.inv_timescales.to(device=hidden_states.device)
        pos_embed = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=-1)
        return pos_embed.to(dtype=hidden_states.dtype)


class Gemma4AudioAttention(nn.Module):
    """Chunked local attention with relative position bias"""

    def __init__(self, config: Gemma4AudioConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.attention_logits_soft_cap = config.attention_logit_cap
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.num_heads = config.num_attention_heads

        self.q_scale = (self.head_dim**-0.5) / math.log(2)
        self.k_scale = math.log(1 + math.e) / math.log(2)

        self.chunk_size = config.attention_chunk_size
        self.max_past_horizon = config.attention_context_left - 1
        self.max_future_horizon = config.attention_context_right
        self.context_size = self.chunk_size + self.max_past_horizon + self.max_future_horizon

        self.q_proj = Gemma4ClippableLinear(config, config.hidden_size, self.num_heads * self.head_dim)
        self.k_proj = Gemma4ClippableLinear(config, config.hidden_size, self.num_heads * self.head_dim)
        self.v_proj = Gemma4ClippableLinear(config, config.hidden_size, self.num_heads * self.head_dim)
        self.post = Gemma4ClippableLinear(config, config.hidden_size, config.hidden_size)

        self.relative_k_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.per_dim_scale = nn.Parameter(torch.zeros(self.head_dim))

        self.register_buffer("softcap", torch.tensor(self.attention_logits_soft_cap), persistent=False)

    def _convert_to_block(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """Splits a `(batch_size, seq_len, num_heads, head_dim)` tensor into non-overlapping blocks of `chunk_size` along the sequence dim."""
        batch_size, seq_len, num_heads, head_dim = hidden_states.shape
        num_blocks = (seq_len + self.chunk_size - 1) // self.chunk_size
        pad = num_blocks * self.chunk_size - seq_len
        hidden_states = F.pad(hidden_states, (0, 0, 0, 0, 0, pad))
        return hidden_states.reshape(batch_size, num_blocks, self.chunk_size, num_heads, head_dim).contiguous()

    def _extract_block_context(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """Extracts overlapping context windows of `context_size` for every block, strided by `chunk_size`."""
        batch_size, seq_len, num_heads, head_dim = hidden_states.shape
        hidden_states = F.pad(
            hidden_states, (0, 0, 0, 0, self.max_past_horizon, self.max_future_horizon + self.chunk_size - 1)
        )
        hidden_states = hidden_states.unfold(1, self.context_size, self.chunk_size)
        hidden_states = torch.movedim(hidden_states, -1, 2)
        return hidden_states.contiguous()

    def _rel_shift(self, x: torch.Tensor) -> torch.Tensor:
        """Relative position shift for blocked attention. See appendix B of https://huggingface.co/papers/1901.02860."""
        batch_size, num_heads, num_blocks, block_size, position_length = x.shape
        context_size = self.context_size
        x = F.pad(x, (0, context_size + 1 - position_length))
        x = x.view(batch_size, num_heads, num_blocks, block_size * (context_size + 1))
        x = x[..., : block_size * context_size]
        return x.view(batch_size, num_heads, num_blocks, block_size, context_size)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: torch.Tensor,
        attention_mask: torch.BoolTensor | None = None,
    ) -> tuple[torch.Tensor, None]:
        batch_size, seq_length, _ = hidden_states.shape
        hidden_shape = (batch_size, seq_length, self.num_heads, self.head_dim)

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

        query_states = query_states * self.q_scale * F.softplus(self.per_dim_scale)
        key_states = key_states * self.k_scale

        query_states = self._convert_to_block(query_states)
        key_states = self._extract_block_context(key_states)
        value_states = self._extract_block_context(value_states)
        num_blocks = query_states.shape[1]

        relative_key_states = self.relative_k_proj(position_embeddings)
        relative_key_states = relative_key_states.view(-1, self.num_heads, self.head_dim)
        relative_key_states = relative_key_states.to(dtype=query_states.dtype)

        queries = query_states.permute(0, 3, 1, 2, 4)
        matrix_ac = queries @ key_states.permute(0, 3, 1, 4, 2)

        queries_flat = queries.reshape(batch_size, self.num_heads, -1, self.head_dim)
        matrix_bd = queries_flat @ relative_key_states.permute(1, 2, 0)
        matrix_bd = matrix_bd.reshape(batch_size, self.num_heads, num_blocks, self.chunk_size, -1)
        matrix_bd = self._rel_shift(matrix_bd)

        attn_weights = matrix_ac + matrix_bd
        attn_weights = attn_weights / self.softcap
        attn_weights = torch.tanh(attn_weights)
        attn_weights = attn_weights * self.softcap

        if attention_mask is not None:
            attn_weights = attn_weights.masked_fill(
                attention_mask.logical_not(), self.config.attention_invalid_logits_value
            )

        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
        attn_output = attn_weights @ value_states.permute(0, 3, 1, 2, 4)
        attn_output = attn_output.permute(0, 2, 3, 1, 4).reshape(batch_size, num_blocks * self.chunk_size, -1)
        attn_output = attn_output[:, :seq_length].contiguous()
        attn_output = self.post(attn_output.to(dtype=self.post.linear.weight.dtype))

        return attn_output, attn_weights


class Gemma4AudioSubSampleConvProjectionLayer(nn.Module):
    def __init__(self, in_channels, out_channels, norm_eps):
        super().__init__()
        self.conv = nn.Conv2d(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=(3, 3),
            stride=(2, 2),
            padding=1,
            bias=False,
        )
        self.norm = nn.LayerNorm(out_channels, eps=norm_eps, elementwise_affine=True, bias=False)
        self.act = nn.ReLU()

    def forward(self, hidden_states: torch.Tensor, mask: torch.Tensor | None = None):
        if mask is not None:
            mask = mask.to(device=hidden_states.device)
            hidden_states = hidden_states * mask[:, None, :, None]

        hidden_states = self.conv(hidden_states.to(self.conv.weight.dtype))
        hidden_states = self.act(self.norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2).contiguous())

        if mask is not None:
            mask = mask[:, ::2]

        return hidden_states, mask


class Gemma4AudioSubSampleConvProjection(nn.Module):
    def __init__(self, config: Gemma4AudioConfig):
        super().__init__()
        self.layer0 = Gemma4AudioSubSampleConvProjectionLayer(
            in_channels=1,
            out_channels=config.subsampling_conv_channels[0],
            norm_eps=config.rms_norm_eps,
        )
        self.layer1 = Gemma4AudioSubSampleConvProjectionLayer(
            in_channels=config.subsampling_conv_channels[0],
            out_channels=config.subsampling_conv_channels[1],
            norm_eps=config.rms_norm_eps,
        )
        proj_input_dim = (config.subsampling_conv_channels[0] // 4) * config.subsampling_conv_channels[1]
        self.input_proj_linear = nn.Linear(proj_input_dim, config.hidden_size, bias=False)

    def forward(
        self,
        input_features: torch.Tensor,
        input_features_mask: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        hidden_states = input_features.unsqueeze(1)
        hidden_states, mask = self.layer0(hidden_states, input_features_mask)
        hidden_states, mask = self.layer1(hidden_states, mask)

        batch_size, _, seq_len, _ = hidden_states.shape
        hidden_states = hidden_states.permute(0, 2, 3, 1).contiguous().reshape(batch_size, seq_len, -1)
        return self.input_proj_linear(hidden_states), mask


class Gemma4AudioFeedForward(nn.Module):
    def __init__(self, config: Gemma4AudioConfig):
        super().__init__()
        self.config = config

        self.ffw_layer_1 = Gemma4ClippableLinear(config, config.hidden_size, config.hidden_size * 4)
        self.ffw_layer_2 = Gemma4ClippableLinear(config, config.hidden_size * 4, config.hidden_size)

        self.pre_layer_norm = Gemma4RMSNorm(config.hidden_size)
        self.post_layer_norm = Gemma4RMSNorm(config.hidden_size)
        self.act_fn = ACT2FN[config.hidden_act]

        self.gradient_clipping = config.gradient_clipping
        self.post_layer_scale = config.residual_weight

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # This is needed to avoid any underflow/overflow issues when clipping
        gradient_clipping = min(self.gradient_clipping, torch.finfo(self.ffw_layer_1.linear.weight.dtype).max)

        residual = hidden_states
        hidden_states = torch.clamp(hidden_states, -gradient_clipping, gradient_clipping)
        hidden_states = self.pre_layer_norm(hidden_states)

        hidden_states = self.ffw_layer_1(hidden_states)
        hidden_states = self.act_fn(hidden_states)
        hidden_states = self.ffw_layer_2(hidden_states)

        hidden_states = torch.clamp(hidden_states, -gradient_clipping, gradient_clipping)
        hidden_states = self.post_layer_norm(hidden_states)
        hidden_states *= self.post_layer_scale
        hidden_states += residual

        return hidden_states


# TODO: this could be imported from Voxtral realtime
class Gemma4AudioCausalConv1d(nn.Conv1d):
    # def __init__(
    #     self,
    #     in_channels: int,
    #     out_channels: int,
    #     kernel_size: int,
    #     # cache_key: str,
    #     stride: int = 1,
    #     dilation: int = 1,
    #     bias: bool = True,
    # ):
    #     super().__init__(in_channels, out_channels, kernel_size, stride=stride, dilation=dilation, bias=bias)
    # self.cache_key = cache_key

    @cached_property
    def left_pad(self):
        effective_kernel_size = (self.kernel_size[0] - 1) * self.dilation[0] + 1
        return effective_kernel_size - self.stride[0]

    def forward(
        self,
        x: torch.Tensor,
        # padding_cache: VoxtralRealtimeConv1dPaddingCache | None = None,  # TODO: we might want to add a cache?
    ) -> torch.Tensor:
        # if padding_cache is not None:
        #     x = padding_cache.update(x, self.cache_key, self)
        # else:
        #     x = nn.functional.pad(x, (self.left_pad, 0))
        x = nn.functional.pad(x, (self.left_pad, 0))

        return super().forward(x)


class Gemma4AudioLightConv1d(nn.Module):
    def __init__(self, config: Gemma4AudioConfig):
        super().__init__()
        self.config = config

        self.linear_start = Gemma4ClippableLinear(config, config.hidden_size, config.hidden_size * 2)
        self.linear_end = Gemma4ClippableLinear(config, config.hidden_size, config.hidden_size)
        self.depthwise_conv1d = Gemma4AudioCausalConv1d(
            in_channels=config.hidden_size,
            out_channels=config.hidden_size,
            kernel_size=config.conv_kernel_size,
            groups=config.hidden_size,
            bias=False,
        )

        self.pre_layer_norm = Gemma4RMSNorm(config.hidden_size, eps=config.rms_norm_eps, with_scale=True)
        self.conv_norm = Gemma4RMSNorm(config.hidden_size, eps=config.rms_norm_eps, with_scale=True)
        self.act_fn = ACT2FN[config.hidden_act]

        self.gradient_clipping = config.gradient_clipping

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        residual = hidden_states

        hidden_states = self.pre_layer_norm(hidden_states)
        hidden_states = self.linear_start(hidden_states)
        hidden_states = nn.functional.glu(hidden_states, dim=-1)

        hidden_states = self.depthwise_conv1d(hidden_states.transpose(1, 2)).transpose(1, 2)

        # This is needed to avoid any underflow/overflow issues when clipping
        gradient_clipping = min(self.gradient_clipping, torch.finfo(self.linear_start.linear.weight.dtype).max)
        hidden_states = torch.clamp(hidden_states, -gradient_clipping, gradient_clipping)
        hidden_states = self.conv_norm(hidden_states)

        hidden_states = self.act_fn(hidden_states)
        hidden_states = self.linear_end(hidden_states)
        hidden_states += residual
        return hidden_states


class Gemma4AudioLayer(nn.Module):
    def __init__(self, config: Gemma4AudioConfig, layer_idx: int):
        super().__init__()
        self.config = config

        self.feed_forward1 = Gemma4AudioFeedForward(config)
        self.feed_forward2 = Gemma4AudioFeedForward(config)
        self.self_attn = Gemma4AudioAttention(config, layer_idx)
        self.lconv1d = Gemma4AudioLightConv1d(config)

        self.norm_pre_attn = Gemma4RMSNorm(config.hidden_size)
        self.norm_post_attn = Gemma4RMSNorm(config.hidden_size)
        self.norm_out = Gemma4RMSNorm(config.hidden_size)

        self.gradient_clipping = config.gradient_clipping

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.BoolTensor | None,
        position_embeddings: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        # This is needed to avoid any underflow/overflow issues when clipping
        gradient_clipping = min(self.gradient_clipping, torch.finfo(self.norm_pre_attn.weight.dtype).max)

        hidden_states = self.feed_forward1(hidden_states)
        residual = hidden_states

        hidden_states = torch.clamp(hidden_states, -gradient_clipping, gradient_clipping)
        hidden_states = self.norm_pre_attn(hidden_states)

        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
        )

        hidden_states = torch.clamp(hidden_states, -gradient_clipping, gradient_clipping)
        hidden_states = self.norm_post_attn(hidden_states)
        hidden_states += residual

        hidden_states = self.lconv1d(hidden_states)
        hidden_states = self.feed_forward2(hidden_states)

        hidden_states = torch.clamp(hidden_states, -gradient_clipping, gradient_clipping)
        hidden_states = self.norm_out(hidden_states)

        return hidden_states


# ---- Vision Encoder Layers ----


class Gemma4VisionPatchEmbedder(nn.Module):
    def __init__(self, config: Gemma4VisionConfig):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.patch_size = config.patch_size
        self.position_embedding_size = config.position_embedding_size

        self.input_proj = nn.Linear(3 * self.patch_size**2, self.hidden_size, bias=False)
        self.position_embedding_table = nn.Parameter(torch.ones(2, self.position_embedding_size, self.hidden_size))

    def _position_embeddings(self, pixel_position_ids: torch.Tensor, padding_positions: torch.Tensor) -> torch.Tensor:
        """Prepare patch positions map for matmul with positon embedding table."""
        # Expanding and permute patch positions to (batch_size, num_patches, 2, position_embedding_size) for matmul.
        clamped_positions = pixel_position_ids.clamp(min=0)
        one_hot = F.one_hot(clamped_positions, num_classes=self.position_embedding_size)
        one_hot = one_hot.permute(0, 2, 1, 3).to(self.position_embedding_table)
        # Compute positional embeddings and sum across x and y.
        position_embeddings = one_hot @ self.position_embedding_table
        position_embeddings = position_embeddings.sum(dim=1)
        # Zero out embeddings for any padding patches.
        position_embeddings = torch.where(padding_positions.unsqueeze(-1), 0.0, position_embeddings)
        return position_embeddings

    def forward(
        self, pixel_values: torch.Tensor, pixel_position_ids: torch.Tensor, padding_positions: torch.Tensor
    ) -> torch.Tensor:
        # Gemma4 applies no normalization and instead scales in model code
        pixel_values = 2 * (pixel_values - 0.5)
        hidden_states = self.input_proj(pixel_values.to(self.input_proj.weight.dtype))
        position_embeddings = self._position_embeddings(pixel_position_ids, padding_positions)
        return hidden_states + position_embeddings


class Gemma4VisionPooler(nn.Module):
    """Scaling and optional spatial pooling for vision encodings"""

    def __init__(self, config: Gemma4VisionConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.root_hidden_size = self.hidden_size**0.5

    def _avg_pool_by_positions(
        self, hidden_states: torch.Tensor, pixel_position_ids: torch.Tensor, length: int
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        2D spatial pooling according to patch positions.
        Pools the input tokens by averaging patches within a `k^2` grid, where `k` is determined by the ratio between
        input and output lengths
        """
        input_seq_len = hidden_states.shape[1]
        k = int((input_seq_len // length) ** 0.5)
        k_squared = k**2
        if k_squared * length != input_seq_len:
            raise ValueError(
                f"Cannot pool {hidden_states.shape} to {length}: {k=}^2 times {length=} must be {input_seq_len}."
            )

        # Clamp padding positions (which are -1) to 0 so they don't break one_hot.
        # Padding patches have zero hidden states so they contribute nothing to the average.
        clamped_positions = pixel_position_ids.clamp(min=0)
        max_x = clamped_positions[..., 0].max(dim=-1, keepdim=True)[0] + 1
        kernel_idxs = torch.div(clamped_positions, k, rounding_mode="floor")
        kernel_idxs = kernel_idxs[..., 0] + (max_x // k) * kernel_idxs[..., 1]
        weights = F.one_hot(kernel_idxs.long(), length).float() / k_squared
        output = weights.transpose(1, 2) @ hidden_states.float()
        mask = torch.logical_not((weights == 0).all(dim=1))
        return output.to(hidden_states.dtype), mask

    def forward(
        self,
        hidden_states: torch.Tensor,
        pixel_position_ids: torch.Tensor,
        padding_positions: torch.Tensor,
        output_length: int | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if output_length > hidden_states.shape[1]:
            raise ValueError(
                f"Cannot output more soft tokens (requested {output_length}) than there are patches"
                f" ({hidden_states.shape[1]}). Change the value of `num_soft_tokens` when processing."
            )

        hidden_states = hidden_states.masked_fill(padding_positions.unsqueeze(-1), 0.0)

        if hidden_states.shape[1] != output_length:
            hidden_states, padding_positions = self._avg_pool_by_positions(
                hidden_states, pixel_position_ids, output_length
            )

        hidden_states *= self.root_hidden_size
        return hidden_states, padding_positions


class Gemma4VisionMLP(nn.Module):
    def __init__(self, config: Gemma4VisionConfig):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = Gemma4ClippableLinear(config, self.hidden_size, self.intermediate_size)
        self.up_proj = Gemma4ClippableLinear(config, self.hidden_size, self.intermediate_size)
        self.down_proj = Gemma4ClippableLinear(config, self.intermediate_size, self.hidden_size)
        self.act_fn = ACT2FN[config.hidden_activation]

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


class Gemma4VisionRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Gemma4VisionConfig, device=None):
        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, device)

        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)

    @staticmethod
    def compute_default_rope_parameters(
        config: Gemma4VisionConfig | None = None,
        device: torch.device | None = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        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

        # The reference implementation computes RoPE frequencies INDEPENDENTLY
        # for each spatial dimension using the partitioned head_dim (head_dim // ndim),
        # so both x and y dimensions get identical frequency ranges.
        # This is different from splitting the global inv_freq between dimensions.
        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.int64).to(device=device, dtype=torch.float) / spatial_dim)
        )
        return inv_freq, attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"

        # Multidimensional positions: [batch, num_patches, ndim]. Apply rotations to each spatial dim separately
        all_cos, all_sin = [], []
        for i in range(2):
            dim_position_ids = position_ids[:, :, i]
            dim_position_ids_expanded = dim_position_ids[:, None, :].float()

            with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
                freqs = (inv_freq_expanded.float() @ dim_position_ids_expanded.float()).transpose(1, 2)
                emb = torch.cat((freqs, freqs), dim=-1)
                cos = emb.cos() * self.attention_scaling
                sin = emb.sin() * self.attention_scaling
            all_cos.append(cos)
            all_sin.append(sin)

        cos = torch.cat(all_cos, dim=-1).to(dtype=x.dtype)
        sin = torch.cat(all_sin, dim=-1).to(dtype=x.dtype)
        return cos, sin


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(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, unsqueeze_dim: int = 1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        x (`torch.Tensor`): The tensor to embed.
        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)
    return (x * cos) + (rotate_half(x) * sin)


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,
    dropout: float | int = 0.0,
    scaling: float | None = None,
    softcap: float | None = None,
    **kwargs,
) -> tuple[torch.Tensor, torch.Tensor]:
    if scaling is None:
        scaling = module.head_dim**-0.5

    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 softcap is not None:
        attn_weights = attn_weights / softcap
        attn_weights = torch.tanh(attn_weights)
        attn_weights = attn_weights * softcap
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    # upcast attention to fp32
    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


#@torch.compile(mode="reduce-overhead")
def _compiled_flex_attention_inner(
    query: torch.Tensor,
    key_states: torch.Tensor,
    value_states: torch.Tensor,
    block_mask,
    scaling: float,
) -> torch.Tensor:
    out = _torch_flex_attention(
        query, key_states, value_states,
        block_mask=block_mask, scale=scaling,
    )
    return out.transpose(1, 2).contiguous()


def flex_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    dropout: float | int = 0.0,
    scaling: float | None = None,
    softcap: float | None = None,
    **kwargs,
) -> tuple[torch.Tensor, None]:
    if not _FLEX_ATTENTION_AVAILABLE:
        raise RuntimeError("torch flex_attention is not available in this PyTorch build")

    if scaling is None:
        scaling = module.head_dim ** -0.5

    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    batch_size, num_heads, q_len, head_dim = query.shape
    kv_len = key_states.shape[2]

    is_sliding = getattr(module, "is_sliding", False)
    sliding_window = getattr(module, "sliding_window", None)

    if is_sliding and sliding_window is not None:
        _sw = sliding_window

        def mask_mod(b, h, q_idx, kv_idx):
            return (q_idx >= kv_idx) & ((q_idx - kv_idx) < _sw)
    else:

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

    block_mask = _torch_create_block_mask(
        mask_mod,
        B=batch_size,
        H=None,
        Q_LEN=q_len,
        KV_LEN=kv_len,
        device=query.device,
    )

    attn_output = _compiled_flex_attention_inner(
        query, key_states, value_states, block_mask, scaling,
    )

    return attn_output, None


def apply_multidimensional_rope(
    x: torch.Tensor,
    cos: torch.Tensor,
    sin: torch.Tensor,
    position_ids: torch.Tensor,
    unsqueeze_dim: int = 2,
) -> torch.Tensor:
    """Applies multidimensional RoPE to inputs.

    Args:
        x (`torch.Tensor`): The tensor to embed.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            If position_ids.ndim + 2 == x.ndim, then this function passes through to `apply_rotary_pos_emb()`.
            Otherwise, position_ids is used to split the inputs, x, into multiple pieces, where each piece is fed to
            `apply_rotary_pos_emb()`, and then concatenated back together.
        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:
      Tensor of shape [B, L, N, H] with RoPE applied.
    """
    ndim = position_ids.shape[-1]
    num_input_channels = x.shape[-1]
    num_rotated_channels_per_dim = 2 * (num_input_channels // (2 * ndim))

    if num_rotated_channels_per_dim <= 0:
        raise ValueError(
            "Invalid configuration: num_rotated_channels_per_dim must be > 0, got"
            f" {num_rotated_channels_per_dim} (num_input_channels={num_input_channels},"
            f" ndim={ndim})"
        )

    # Correctly split the input tensor into ndim parts
    split_sizes = [num_rotated_channels_per_dim] * ndim
    x_parts = torch.split(x, split_sizes, dim=-1)
    cos_parts = torch.split(cos, split_sizes, dim=-1)
    sin_parts = torch.split(sin, split_sizes, dim=-1)
    y_parts = [
        apply_rotary_pos_emb(
            x=x_parts[k],
            cos=cos_parts[k],
            sin=sin_parts[k],
            unsqueeze_dim=unsqueeze_dim,
        )
        for k in range(ndim)
    ]
    return torch.cat(y_parts, dim=-1)


@use_kernelized_func(apply_rotary_pos_emb)
class Gemma4VisionAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Gemma4VisionConfig, 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 = 1.0
        self.attention_dropout = self.config.attention_dropout
        self.is_causal = False
        self.q_proj = Gemma4ClippableLinear(config, config.hidden_size, config.num_attention_heads * self.head_dim)
        self.k_proj = Gemma4ClippableLinear(config, config.hidden_size, config.num_key_value_heads * self.head_dim)
        self.v_proj = Gemma4ClippableLinear(config, config.hidden_size, config.num_key_value_heads * self.head_dim)
        self.o_proj = Gemma4ClippableLinear(config, config.num_attention_heads * self.head_dim, config.hidden_size)

        self.q_norm = Gemma4RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps)
        self.k_norm = Gemma4RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps)
        self.v_norm = Gemma4RMSNorm(self.head_dim, eps=config.rms_norm_eps, with_scale=False)

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

        cos, sin = position_embeddings

        query_states = self.q_proj(hidden_states).view(hidden_shape)
        query_states = self.q_norm(query_states)
        query_states = apply_multidimensional_rope(query_states, cos, sin, position_ids)
        query_states = query_states.transpose(1, 2)

        key_states = self.k_proj(hidden_states).view(hidden_shape)
        key_states = self.k_norm(key_states)
        key_states = apply_multidimensional_rope(key_states, cos, sin, position_ids)
        key_states = key_states.transpose(1, 2)

        value_states = self.v_proj(hidden_states).view(hidden_shape)
        value_states = self.v_norm(value_states)
        value_states = value_states.transpose(1, 2)

        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=self.attention_dropout if self.training else 0.0,
            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 Gemma4VisionEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Gemma4VisionConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.self_attn = Gemma4VisionAttention(config=config, layer_idx=layer_idx)
        self.mlp = Gemma4VisionMLP(config)
        self.input_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.pre_feedforward_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.post_feedforward_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: torch.Tensor = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            **kwargs,
        )
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.pre_feedforward_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states


class Gemma4VisionEncoder(nn.Module):
    def __init__(self, config: Gemma4VisionConfig):
        super().__init__()
        self.config = config
        self.num_layers = config.num_hidden_layers
        self.rotary_emb = Gemma4VisionRotaryEmbedding(config)
        self.layers = nn.ModuleList(
            [Gemma4VisionEncoderLayer(config=config, layer_idx=i) for i in range(self.num_layers)]
        )

    def forward(
        self,
        inputs_embeds: torch.Tensor,
        attention_mask: torch.Tensor,
        pixel_position_ids: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        r"""
        pixel_position_ids (torch.Tensor):
            Patch positions as (x, y) coordinates in the image as [batch, num_patches, 2].
        """
        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
        )

        # embed positions
        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, pixel_position_ids)

        # decoder layers
        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_embeddings=position_embeddings,
                position_ids=pixel_position_ids,
                **kwargs,
            )

        return BaseModelOutputWithPast(last_hidden_state=hidden_states)


class Gemma4TextMLP(nn.Module):
    def __init__(self, config: Gemma4TextConfig, layer_idx: int):
        super().__init__()
        first_kv_shared_layer_idx = config.num_hidden_layers - config.num_kv_shared_layers
        is_kv_shared_layer = layer_idx >= first_kv_shared_layer_idx > 0
        use_double_wide_mlp = config.use_double_wide_mlp and is_kv_shared_layer
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size * (2 if use_double_wide_mlp else 1)
        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_activation]

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


class Gemma4TextRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Gemma4TextConfig, device=None, layer_type=None):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.layer_types = set(config.layer_types)
        self.rope_init_fns: dict[str, Callable[..., tuple[torch.Tensor, float]]] = {}
        self.rope_type: dict[str, str] = {}

        for layer_type in self.layer_types:
            rope_params = self.config.rope_parameters[layer_type]
            if rope_params is None:
                continue

            if (rope_type := rope_params["rope_type"]) != "default":
                rope_init_fn = ROPE_INIT_FUNCTIONS[rope_type]
            else:
                rope_init_fn = self.compute_default_rope_parameters

            self.rope_init_fns[layer_type] = rope_init_fn
            self.rope_type[layer_type] = rope_type

            rope_init_fn_kwargs = {"device": device, "layer_type": layer_type}
            if layer_type == "full_attention" and rope_type == "proportional":
                rope_init_fn_kwargs["head_dim_key"] = "global_head_dim"

            curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, **rope_init_fn_kwargs)
            self.register_buffer(f"{layer_type}_inv_freq", curr_inv_freq, persistent=False)
            self.register_buffer(f"{layer_type}_original_inv_freq", curr_inv_freq.clone(), persistent=False)
            setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)

    @staticmethod
    def compute_default_rope_parameters(
        config: Gemma4TextConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
        layer_type: str | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
            layer_type (`str`, *optional*):
                The current layer type if the model has different RoPE parameters per type.
                Should not be used unless `config.layer_types is not None`

        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).
        """
        # For backward compatibility standardize the `rope_parameters_dict` if it uses old format
        base = config.rope_parameters[layer_type]["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.int64).to(device=device, dtype=torch.float) / dim)
        )
        return inv_freq, attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids, layer_type=None):
        inv_freq = getattr(self, f"{layer_type}_inv_freq")
        attention_scaling = getattr(self, f"{layer_type}_attention_scaling")

        inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * attention_scaling
            sin = emb.sin() * attention_scaling

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


@use_kernelized_func(apply_rotary_pos_emb)
class Gemma4TextAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Gemma4TextConfig, 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.is_sliding = self.layer_type == "sliding_attention"
        self.sliding_window = config.sliding_window if self.is_sliding else None

        self.head_dim = config.global_head_dim if not self.is_sliding and config.global_head_dim else config.head_dim
        self.use_alternative_attention = config.attention_k_eq_v and not self.is_sliding
        num_key_value_heads = (
            config.num_global_key_value_heads if self.use_alternative_attention else config.num_key_value_heads
        )
        self.num_key_value_groups = config.num_attention_heads // num_key_value_heads
        self.scaling = 1.0
        self.attention_dropout = self.config.attention_dropout
        self.is_causal = config.use_bidirectional_attention != "all"

        # Shared kv cache
        first_kv_shared_layer_idx = self.config.num_hidden_layers - getattr(self.config, "num_kv_shared_layers", 0)
        self.is_kv_shared_layer = layer_idx >= first_kv_shared_layer_idx > 0
        prev_layers = config.layer_types[:first_kv_shared_layer_idx]
        if self.is_kv_shared_layer:
            # For shared layers, find the last non-shared layer of the same type before sharing starts
            self.kv_shared_layer_index = len(prev_layers) - 1 - prev_layers[::-1].index(config.layer_types[layer_idx])
            self.store_full_length_kv = False
        else:
            self.kv_shared_layer_index = None
            # For non-shared layers, store full-length kv if this is the last non-shared layer of its type
            self.store_full_length_kv = layer_idx == len(prev_layers) - 1 - prev_layers[::-1].index(
                config.layer_types[layer_idx]
            )

        self.q_norm = Gemma4RMSNorm(dim=self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = Gemma4RMSNorm(dim=self.head_dim, eps=config.rms_norm_eps)
        self.v_norm = Gemma4RMSNorm(self.head_dim, eps=config.rms_norm_eps, with_scale=False)

        self.k_proj = nn.Linear(config.hidden_size, num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = (
            nn.Linear(config.hidden_size, num_key_value_heads * self.head_dim, bias=config.attention_bias)
            if not self.use_alternative_attention
            else None
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

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

        cos, sin = position_embeddings

        query_states = self.q_proj(hidden_states).view(hidden_shape)
        query_states = self.q_norm(query_states)
        query_states = apply_rotary_pos_emb(query_states, cos, sin, unsqueeze_dim=2)
        query_states = query_states.transpose(1, 2)

        # For layers with shared KV (from kv sharing point onwards), we reuse the same keys/values states as the last non-sharing layer
        if self.is_kv_shared_layer and past_key_values is not None:
            key_states, value_states = past_key_values.shared_layers[self.kv_shared_layer_index]
            # Device of past layer may be different from current one
            key_states = key_states.to(query_states.device)
            value_states = value_states.to(query_states.device)
        else:
            key_states = self.k_proj(hidden_states).view(hidden_shape)
            value_states = self.v_proj(hidden_states).view(hidden_shape) if self.v_proj is not None else key_states

            key_states = self.k_norm(key_states)
            key_states = apply_rotary_pos_emb(key_states, cos, sin, unsqueeze_dim=2)
            key_states = key_states.transpose(1, 2)

            value_states = self.v_norm(value_states)
            value_states = value_states.transpose(1, 2)

        if past_key_values is not None:
            if not self.is_kv_shared_layer:
                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
            if self.store_full_length_kv:
                if not hasattr(past_key_values, "shared_layers"):
                    past_key_values.shared_layers = {}
                past_key_values.shared_layers[self.layer_idx] = key_states, value_states

        # --- Attention dispatch: eager / flex / hybrid / verify ---
        #   "eager"  : always eager  (default, safe)
        #   "flex"   : flex for ALL layers (only useful for very long SWA)
        #   "hybrid" : flex for SWA layers only, eager for Full  (recommended)
        #   "verify" : run both per layer and log differences
        mode = _FLEX_ATTENTION_MODE
        q_len = query_states.shape[2]
        _use_flex = (
            _FLEX_ATTENTION_AVAILABLE
            and q_len > 1
            and (
                mode == "flex"
                or (mode == "hybrid" and self.is_sliding)
            )
        )
        _attn_kwargs = dict(
            dropout=self.attention_dropout if self.training else 0.0,
            scaling=self.scaling,
            sliding_window=self.sliding_window,
            **kwargs,
        )

        if _use_flex:
            attn_output, attn_weights = flex_attention_forward(
                self, query_states, key_states, value_states, attention_mask,
                **_attn_kwargs,
            )
        elif mode == "verify" and _FLEX_ATTENTION_AVAILABLE and q_len > 1:
            eager_output, eager_weights = eager_attention_forward(
                self, query_states, key_states, value_states, attention_mask,
                **_attn_kwargs,
            )
            flex_output, _ = flex_attention_forward(
                self, query_states, key_states, value_states, attention_mask,
                **_attn_kwargs,
            )
            _e_f = eager_output.float()
            _f_f = flex_output.float()
            _max_diff = (_e_f - _f_f).abs().max().item()
            _cos = F.cosine_similarity(
                _e_f.flatten().unsqueeze(0), _f_f.flatten().unsqueeze(0)
            ).item()
            print(
                f"[VERIFY] Layer {self.layer_idx:2d} "
                f"({self.layer_type:20s}): "
                f"max_diff={_max_diff:.4e}  cos_sim={_cos:.6f}"
            )
            attn_output = eager_output
            attn_weights = eager_weights
        else:
            attn_output, attn_weights = eager_attention_forward(
                self, query_states, key_states, value_states, attention_mask,
                **_attn_kwargs,
            )

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

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights


@use_experts_implementation
class Gemma4TextExperts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config: Gemma4TextConfig):
        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_activation]

    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)
            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 Gemma4TextRouter(nn.Module):
    def __init__(self, config: Gemma4TextConfig):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.scalar_root_size = self.hidden_size**-0.5
        self.eps = config.rms_norm_eps

        self.norm = Gemma4RMSNorm(self.hidden_size, eps=self.eps, with_scale=False)
        self.proj = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.scale = nn.Parameter(torch.ones(self.hidden_size))
        self.per_expert_scale = nn.Parameter(torch.ones(config.num_experts))

    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        hidden_states = self.norm(hidden_states)
        hidden_states = hidden_states * self.scale * self.scalar_root_size

        expert_scores = self.proj(hidden_states)  # [B*S, E]
        router_probabilities = nn.functional.softmax(expert_scores, dim=-1)

        # topk returns both values (probabilities) and indices directly
        top_k_weights, top_k_index = torch.topk(
            router_probabilities,
            k=self.config.top_k_experts,
            dim=-1,
        )  # both [B*S, K]

        # Normalize the top-k weights so they sum to 1 per token
        top_k_weights /= top_k_weights.sum(dim=-1, keepdim=True)

        # Apply per-expert scale directly to the weights
        top_k_weights = top_k_weights * self.per_expert_scale[top_k_index]

        return router_probabilities, top_k_weights, top_k_index


class Gemma4TextDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Gemma4TextConfig | Gemma4VisionConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.self_attn = Gemma4TextAttention(config=config, layer_idx=layer_idx)
        self.mlp = Gemma4TextMLP(config, layer_idx)
        self.input_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.pre_feedforward_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.post_feedforward_layernorm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.register_buffer("layer_scalar", torch.ones(1))

        self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
        if self.hidden_size_per_layer_input:
            self.act_fn = ACT2FN[config.hidden_activation]
            self.per_layer_input_gate = nn.Linear(self.hidden_size, self.hidden_size_per_layer_input, bias=False)
            self.per_layer_projection = nn.Linear(self.hidden_size_per_layer_input, self.hidden_size, bias=False)
            self.post_per_layer_input_norm = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)

        self.enable_moe_block = config.enable_moe_block
        if self.enable_moe_block:
            self.router = Gemma4TextRouter(config)
            self.experts = Gemma4TextExperts(config)
            self.post_feedforward_layernorm_1 = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
            self.post_feedforward_layernorm_2 = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
            self.pre_feedforward_layernorm_2 = Gemma4RMSNorm(self.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        per_layer_input: torch.Tensor = None,
        position_embeddings: torch.Tensor = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        output_attentions: bool = False,
        **kwargs,
    ) -> tuple[torch.Tensor, ...]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, attn_output = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            **kwargs,
        )
        #attn_output = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.pre_feedforward_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)

        if self.enable_moe_block:
            hidden_states_1 = self.post_feedforward_layernorm_1(hidden_states)

            # Take hidden states before MLP here
            hidden_states_flat = residual.reshape(-1, residual.shape[-1])
            _, top_k_weights, top_k_index = self.router(hidden_states_flat)
            hidden_states_2 = self.pre_feedforward_layernorm_2(hidden_states_flat)
            hidden_states_2 = self.experts(hidden_states_2, top_k_index, top_k_weights)
            hidden_states_2 = hidden_states_2.reshape(residual.shape)
            hidden_states_2 = self.post_feedforward_layernorm_2(hidden_states_2)

            # Combine mlp and moe outputs
            hidden_states = hidden_states_1 + hidden_states_2

        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        if self.hidden_size_per_layer_input:
            residual = hidden_states
            hidden_states = self.per_layer_input_gate(hidden_states)
            hidden_states = self.act_fn(hidden_states)
            hidden_states = hidden_states * per_layer_input
            hidden_states = self.per_layer_projection(hidden_states)
            hidden_states = self.post_per_layer_input_norm(hidden_states)
            hidden_states = residual + hidden_states

        hidden_states *= self.layer_scalar

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (attn_output,)

        return outputs


class Gemma4TextScaledWordEmbedding(nn.Embedding):
    """
    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float = 1.0):
        super().__init__(num_embeddings, embedding_dim, padding_idx)
        self.scalar_embed_scale = embed_scale
        self.register_buffer("embed_scale", torch.tensor(embed_scale), persistent=False)

    def forward(self, input_ids: torch.Tensor):
        return super().forward(input_ids) * self.embed_scale.to(self.weight.dtype)


# ---- Model Classes ----


class Gemma4PreTrainedModel(PreTrainedModel):
    config: Gemma4Config
    supports_gradient_checkpointing = True
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _no_split_modules = ["Gemma4TextDecoderLayer", "Gemma4VisionEncoderLayer", "Gemma4AudioLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    input_modalities = ("image", "text", "video", "audio")

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Gemma4VisionPatchEmbedder):
            init.ones_(module.position_embedding_table)
        elif isinstance(module, Gemma4AudioRelPositionalEncoding):
            min_timescale = 1.0
            max_timescale = 10000.0
            num_timescales = module.hidden_size // 2
            log_timescale_increment = math.log(max_timescale / min_timescale) / max(num_timescales - 1, 1)
            inv_timescales = min_timescale * torch.exp(torch.arange(num_timescales) * -log_timescale_increment)
            init.copy_(module.inv_timescales, inv_timescales.unsqueeze(0).unsqueeze(0))
        elif isinstance(module, Gemma4AudioAttention):
            init.constant_(module.softcap, module.attention_logits_soft_cap)
            init.zeros_(module.per_dim_scale)
        elif isinstance(module, Gemma4TextRotaryEmbedding):
            for layer_type, rope_init_fn in module.rope_init_fns.items():
                rope_init_fn_kwargs = {"layer_type": layer_type}
                if layer_type == "full_attention" and module.rope_type[layer_type] == "proportional":
                    rope_init_fn_kwargs["head_dim_key"] = "global_head_dim"

                curr_inv_freq, _ = rope_init_fn(module.config, **rope_init_fn_kwargs)
                init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
                init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)
        elif isinstance(module, Gemma4VisionRotaryEmbedding):
            rope_fn = (
                ROPE_INIT_FUNCTIONS[module.rope_type]
                if module.rope_type != "default"
                else module.compute_default_rope_parameters
            )
            buffer_value, _ = rope_fn(module.config)
            init.copy_(module.inv_freq, buffer_value)
            init.copy_(module.original_inv_freq, buffer_value)
        elif isinstance(module, Gemma4TextScaledWordEmbedding):
            init.constant_(module.embed_scale, module.scalar_embed_scale)
        elif isinstance(module, Gemma4TextRouter):
            init.ones_(module.scale)
            init.ones_(module.per_expert_scale)
        elif isinstance(module, Gemma4TextExperts):
            std = self.config.initializer_range
            init.normal_(module.gate_up_proj, mean=0.0, std=std)
            init.normal_(module.down_proj, mean=0.0, std=std)
        elif isinstance(module, Gemma4TextDecoderLayer):
            init.ones_(module.layer_scalar)
        elif isinstance(module, Gemma4ClippableLinear) and module.use_clipped_linears:
            init.constant_(module.input_min, -float("inf"))
            init.constant_(module.input_max, float("inf"))
            init.constant_(module.output_min, -float("inf"))
            init.constant_(module.output_max, float("inf"))
        elif isinstance(module, Gemma4VisionModel) and module.config.standardize:
            init.zeros_(module.std_bias)
            init.ones_(module.std_scale)


@auto_docstring(custom_intro="The base Gemma 4 language model without a language modeling head.")
class Gemma4TextModel(Gemma4PreTrainedModel):
    config: Gemma4TextConfig
    input_modalities = ("text",)
    _can_record_outputs = {
        "router_logits": OutputRecorder(Gemma4TextRouter, index=0),
        "hidden_states": Gemma4TextDecoderLayer,
    }

    def __init__(self, config: Gemma4TextConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        # Gemma4 downcasts the below to bfloat16, causing sqrt(3072)=55.4256 to become 55.5. See https://github.com/huggingface/transformers/pull/29402
        self.embed_tokens = Gemma4TextScaledWordEmbedding(
            config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=self.config.hidden_size**0.5
        )
        self.layers = nn.ModuleList(
            [Gemma4TextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Gemma4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Gemma4TextRotaryEmbedding(config)
        self.gradient_checkpointing = False
        self.unique_layer_types = set(self.config.layer_types)

        self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
        if self.hidden_size_per_layer_input:
            self.embed_tokens_per_layer = Gemma4TextScaledWordEmbedding(
                config.vocab_size_per_layer_input,
                config.num_hidden_layers * config.hidden_size_per_layer_input,
                self.padding_idx,
                embed_scale=config.hidden_size_per_layer_input**0.5,
            )
            self.per_layer_input_scale = 2.0**-0.5
            self.per_layer_model_projection = nn.Linear(
                config.hidden_size,
                config.num_hidden_layers * config.hidden_size_per_layer_input,
                bias=False,
            )
            self.per_layer_model_projection_scale = config.hidden_size**-0.5
            self.per_layer_projection_norm = Gemma4RMSNorm(config.hidden_size_per_layer_input, eps=config.rms_norm_eps)

        # 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.FloatTensor | None = None,
        per_layer_inputs: torch.Tensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        r"""
        per_layer_inputs (`torch.Tensor` of shape `(batch_size, sequence_length, num_hidden_layers, hidden_size_per_layer_input)`, *optional*):
            Pre-computed per-layer input embeddings. When provided, these are used directly instead of being
            computed from `input_ids` via `get_per_layer_inputs()`. This is primarily used by the multimodal
            model (`Gemma4Model`) which pre-computes per-layer inputs from the original `input_ids` *before*
            merging multimodal soft tokens into `inputs_embeds` — at which point the original token ids are
            no longer recoverable.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

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

        if input_ids is not None:
            inputs_embeds = self.embed_tokens(input_ids)

        if self.hidden_size_per_layer_input:
            if per_layer_inputs is None:
                per_layer_inputs = self.get_per_layer_inputs(input_ids, inputs_embeds)
            per_layer_inputs = self.project_per_layer_inputs(inputs_embeds, per_layer_inputs)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        # KV shared layers need a cache object to store/retrieve shared KV states,
        # even when use_cache=False. Without this, shared layers fall back to
        # computing their own K/V projections instead of reusing from earlier layers.
        _kv_sharing_only = False
        if past_key_values is None and getattr(self.config, "num_kv_shared_layers", 0) > 0:
            past_key_values = DynamicCache(config=self.config)
            _kv_sharing_only = not use_cache

        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.unsqueeze(0)

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # --- PATCH: force attn_impl to "eager" for mask creation so that
            # create_causal_mask materializes a real 4D causal mask. Without
            # this, HF returns None when _attn_implementation is "sdpa"/"flash_attention_2"
            # (those handle causality internally), but our attention path
            # always uses eager_attention_forward which does NOT.  Missing
            # causal mask => full_attention layers become non-causal, i.e.
            # future tokens leak into past predictions.
            _saved_impl = self.config._attn_implementation
            self.config._attn_implementation = "eager"
            try:
                # Prepare mask arguments
                mask_kwargs = {
                    "config": self.config,
                    "inputs_embeds": inputs_embeds,
                    "attention_mask": attention_mask,
                    "past_key_values": past_key_values,
                    "position_ids": position_ids,
                }
                # Create the masks
                causal_mask_mapping = {
                    "full_attention": create_causal_mask(**mask_kwargs),
                    "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),
                }
            finally:
                self.config._attn_implementation = _saved_impl

        # embed positions
        hidden_states = inputs_embeds
        position_embeddings = {}
        for layer_type in self.unique_layer_types:
            position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            per_layer_input = per_layer_inputs[:, :, i, :] if per_layer_inputs is not None else None

            layer_outputs = decoder_layer(
                hidden_states,
                per_layer_input,
                position_embeddings=position_embeddings[self.config.layer_types[i]],
                attention_mask=causal_mask_mapping[self.config.layer_types[i]],
                position_ids=position_ids,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                **kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if not _kv_sharing_only else None,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )

    def get_per_layer_inputs(self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None) -> torch.Tensor:
        if not self.hidden_size_per_layer_input:
            raise RuntimeError(
                "Attempting to call get_per_layer_inputs() from a model initialized with a config that does not support"
                f" per-layer embeddings. {self.config}"
            )

        # If only inputs_embeds are provided, reverse main embedding to find the input_ids - this allows to `generate`
        # from `inputs_embeds` only as other models (otherwise it would need the value from both embeddings)
        if input_ids is None:
            with torch.no_grad():
                input_ids = (
                    (
                        inputs_embeds[:, :, None, :]
                        == self.embed_tokens.weight[None, None, :, :] * self.config.hidden_size**0.5
                    )
                    .all(dim=3)
                    .nonzero()[:, 2]
                )
                try:
                    input_ids = input_ids.view(inputs_embeds.shape[:2])
                except RuntimeError:
                    raise RuntimeError(
                        "It seems like you tried to call `forward` from `inputs_embeds` without providing `input_ids`, and that "
                        "the `inputs_embeds` you provided do not exactly match the embedding weights. Since Gemma4 needs to reverse "
                        "the embedding to compute another embedding, make sure you provide exact `inputs_embeds`"
                    )

        return self.embed_tokens_per_layer(input_ids).reshape(
            *input_ids.shape,
            self.config.num_hidden_layers,
            self.hidden_size_per_layer_input,
        )

    def project_per_layer_inputs(
        self,
        inputs_embeds: torch.Tensor,
        per_layer_inputs: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if not self.hidden_size_per_layer_input:
            raise RuntimeError(
                "Attempting to call project_per_layer_inputs() from a model initialized with a config that does not"
                f" support per-layer embeddings. {self.config}"
            )

        per_layer_projection = self.per_layer_model_projection(inputs_embeds) * self.per_layer_model_projection_scale
        per_layer_projection = per_layer_projection.reshape(
            *inputs_embeds.shape[:-1],
            self.config.num_hidden_layers,
            self.hidden_size_per_layer_input,
        )
        per_layer_projection = self.per_layer_projection_norm(per_layer_projection)

        if per_layer_inputs is None:
            return per_layer_projection

        return (per_layer_projection + per_layer_inputs) * self.per_layer_input_scale


@auto_docstring(custom_intro="The base Gemma 4 language model with a language modeling head.")
class Gemma4ForCausalLM(Gemma4PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    config: Gemma4TextConfig
    base_model_prefix = "model"

    def __init__(self, config: Gemma4TextConfig):
        super().__init__(config)
        self.model = Gemma4TextModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    @can_return_tuple
    @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.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from transformers import AutoTokenizer, Gemma4ForCausalLM

        >>> model = Gemma4ForCausalLM.from_pretrained("google/gemma-2-9b")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")

        >>> prompt = "What is your favorite condiment?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "What is your favorite condiment?"
        ```"""
        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs: BaseModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        if self.config.final_logit_softcapping is not None:
            logits = logits / self.config.final_logit_softcapping
            logits = torch.tanh(logits)
            logits = logits * self.config.final_logit_softcapping

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

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


def sliding_window_mask_function(sliding_window: tuple[int, int]) -> Callable:
    """
    This creates uni/bidirectional attention mask with sliding window.
    """

    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
        left_window_size, right_window_size = sliding_window

        dist = q_idx - kv_idx
        left_mask = (dist >= 0) & (dist < left_window_size)
        right_mask = (dist < 0) & (-dist < right_window_size)
        return left_mask | right_mask

    return inner_mask


class Gemma4AudioModel(Gemma4PreTrainedModel):
    """An audio encoder based on the [Universal Speech Model](https://huggingface.co/papers/2303.01037) architecture."""

    config: Gemma4AudioConfig
    main_input_name = "input_features"
    base_model_prefix = "model.audio_tower"  # prefix for Gemma4ForConditionalGeneration saved checkpoints, required for Gemma4AudioModel.from_pretrained()
    _can_record_outputs = {
        "hidden_states": Gemma4AudioLayer,
        "attentions": Gemma4AudioAttention,
    }

    def __init__(self, config: Gemma4AudioConfig):
        super().__init__(config)
        self.config = config

        self.subsample_conv_projection = Gemma4AudioSubSampleConvProjection(config)
        self.rel_pos_enc = Gemma4AudioRelPositionalEncoding(config)
        self.layers = nn.ModuleList(
            [Gemma4AudioLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.output_proj = nn.Linear(config.hidden_size, config.output_proj_dims, bias=True)

        self.post_init()

    def _convert_4d_mask_to_blocked_5d(self, mask_4d: torch.Tensor) -> torch.Tensor:
        """
        Convert a standard 4D attention mask `[batch_size, 1, seq_len, seq_len]` to the 5D blocked format
        `[batch_size, 1, num_blocks, chunk_size, context_size]` expected by the chunked local attention,
        """
        batch_size, _, seq_len, _ = mask_4d.shape
        device = mask_4d.device

        chunk_size = self.config.attention_chunk_size
        max_past_horizon = self.config.attention_context_left - 1
        max_future_horizon = self.config.attention_context_right

        num_blocks = (seq_len + chunk_size - 1) // chunk_size
        padded_seq_len = num_blocks * chunk_size
        pad_amount = padded_seq_len - seq_len

        mask_4d = F.pad(mask_4d, (0, pad_amount, 0, pad_amount), value=False)
        mask_5d = mask_4d.reshape(batch_size, 1, num_blocks, chunk_size, padded_seq_len)
        mask_5d = F.pad(mask_5d, (max_past_horizon, max_future_horizon), value=False)

        block_starts = torch.arange(num_blocks, device=device) * chunk_size
        offsets = torch.arange(chunk_size + max_past_horizon + max_future_horizon, device=device)
        kv_indices = block_starts[:, None] + offsets[None, :]
        kv_indices = kv_indices[None, None, :, None, :].expand(batch_size, 1, -1, chunk_size, -1)

        return mask_5d.gather(-1, kv_indices)

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring(custom_intro="Encodes audio features to soft tokens.")
    def forward(
        self,
        input_features: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.BoolTensor]:
        hidden_states, output_mask = self.subsample_conv_projection(input_features, attention_mask)
        position_embeddings = self.rel_pos_enc(hidden_states)

        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=hidden_states,
            attention_mask=output_mask,
            and_mask_function=sliding_window_mask_function(
                (self.config.attention_context_left - 1, self.config.attention_context_right)
            ),
        )
        attention_mask = self._convert_4d_mask_to_blocked_5d(attention_mask)

        for encoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = encoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.output_proj(hidden_states)
        return Gemma4AudioModelOutput(last_hidden_state=hidden_states, attention_mask=output_mask)


class Gemma4VisionModel(Gemma4PreTrainedModel):
    """The Gemma 4 Vision Encoder."""

    config = Gemma4VisionConfig
    _can_record_outputs = {
        "hidden_states": Gemma4VisionEncoderLayer,
        "attentions": Gemma4VisionAttention,
    }

    def __init__(self, config: Gemma4VisionConfig):
        super().__init__(config)
        self.patch_embedder = Gemma4VisionPatchEmbedder(config)
        self.encoder = Gemma4VisionEncoder(config)
        self.pooler = Gemma4VisionPooler(config)

        if self.config.standardize:
            self.register_buffer("std_bias", torch.empty(self.config.hidden_size))
            self.register_buffer("std_scale", torch.empty(self.config.hidden_size))

        self.post_init()

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring(custom_intro="Encodes image pixels to soft tokens from patches.")
    def forward(
        self,
        pixel_values: torch.FloatTensor,
        pixel_position_ids: torch.LongTensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        r"""
        pixel_values (`torch.FloatTensor` or `list[torch.FloatTensor]`):
            The images to encode. Either a single `[batch, channels, height, width]` tensor
            (all images same size) or a list of `[1, channels, height, width]` tensors (different sizes).
        pixel_position_ids (`torch.LongTensor` of shape `(batch_size, max_patches, 2)`):
            The patch positions as (x, y) coordinates in the image. Padding patches are indicated by (-1, -1).
        """
        pooling_kernel_size = self.config.pooling_kernel_size
        output_length = pixel_values.shape[-2] // (pooling_kernel_size * pooling_kernel_size)

        padding_positions = (pixel_position_ids == -1).all(dim=-1)
        inputs_embeds = self.patch_embedder(pixel_values, pixel_position_ids, padding_positions)
        output = self.encoder(
            inputs_embeds=inputs_embeds,
            attention_mask=~padding_positions,  # encoder expects True=valid, padding_positions is True=padding
            pixel_position_ids=pixel_position_ids,
            **kwargs,
        )

        hidden_states, pooler_mask = self.pooler(
            hidden_states=output.last_hidden_state,
            pixel_position_ids=pixel_position_ids,
            padding_positions=padding_positions,
            output_length=output_length,
        )

        # Strip padding tokens. pooler_mask is True = valid, False = padding.
        hidden_states = hidden_states[pooler_mask]

        if self.config.standardize:
            hidden_states = (hidden_states - self.std_bias) * self.std_scale

        return BaseModelOutputWithPast(last_hidden_state=hidden_states)


class Gemma4MultimodalEmbedder(nn.Module):
    """Embeds token ids or soft tokens for multimodal content into language model space."""

    def __init__(
        self,
        multimodal_config: Gemma4AudioConfig | Gemma4VisionConfig,
        text_config: Gemma4TextConfig,
    ):
        super().__init__()

        self.multimodal_hidden_size = getattr(multimodal_config, "output_proj_dims", multimodal_config.hidden_size)
        self.eps = multimodal_config.rms_norm_eps
        self.text_hidden_size = text_config.hidden_size
        self.embedding_projection = nn.Linear(self.multimodal_hidden_size, self.text_hidden_size, bias=False)
        self.embedding_pre_projection_norm = Gemma4RMSNorm(self.multimodal_hidden_size, eps=self.eps, with_scale=False)

    def forward(self, inputs_embeds: torch.Tensor) -> torch.Tensor:
        """Embeds token ids or soft tokens for multimodal content into language model space.
        Args:
            inputs_embeds: A torch.Tensor containing the soft tokens to embed.
        Returns:
            A torch.Tensor of embeddings with shape `[batch_size, seq_len, self.config.text_config.hidden_size]`.
        """
        embs_normed = self.embedding_pre_projection_norm(inputs_embeds)
        return self.embedding_projection(embs_normed)


# Identical as Gemma3 but modular can't resolve if we simply import. FIXME: @cyril
def token_type_ids_mask_function(
    token_type_ids: torch.Tensor | None,
    image_group_ids: torch.Tensor | None,
) -> Callable | None:
    """
    This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths,
    not start and end indices.
    """
    # Do not return an additional mask in this case
    if token_type_ids is None:
        return None

    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
        seq_length = image_group_ids.shape[-1]

        # clamp indices because with static cache they can go beyond `image_group_ids.shape[-1]`
        q_idx_clamped = q_idx.clamp(max=seq_length - 1)
        kv_idx_clamped = kv_idx.clamp(max=seq_length - 1)

        # Unmask if the q and kv come from same group which is not -1 (i.e. non-text)
        q_group = image_group_ids[batch_idx, q_idx_clamped]
        kv_group = image_group_ids[batch_idx, kv_idx_clamped]
        q_group = torch.where(q_idx < seq_length, q_group, -1)
        kv_group = torch.where(kv_idx < seq_length, kv_group, -1)
        return (q_group == kv_group) & (q_group >= 0)

    return inner_mask


# Similar to Gemma3 but `sliding_mask_kwargs` and `mask_kwargs` are different and `token_type_ids->mm_token_type_ids`
def create_causal_mask_mapping(
    config: PreTrainedConfig,
    inputs_embeds: torch.Tensor,
    attention_mask: torch.Tensor | None,
    past_key_values: Cache | None,
    position_ids: torch.Tensor | None,
    mm_token_type_ids: torch.Tensor | None = None,
    pixel_values: torch.FloatTensor | None = None,
    is_training: bool = False,
    is_first_iteration: bool | None = None,
    **kwargs,
) -> dict:
    """
    Overwrites the base `create_masks_for_generate` with `token_type_ids` masking to create the causal mask mapping
    for all kinds of forward passes. Gemma4 uses a bidirectional mask for images.

    Uses `pixel_values` as an optional input to disambiguate edge cases.
    """
    if is_training and mm_token_type_ids is None:
        raise ValueError("`mm_token_type_ids` is required as a model input when training")

    mask_kwargs = {
        "config": config.get_text_config(),
        "inputs_embeds": inputs_embeds,
        "attention_mask": attention_mask,
        "past_key_values": past_key_values,
        "position_ids": position_ids,
    }
    sliding_mask_kwargs = mask_kwargs.copy()

    # NOTE: this `may_have_image_input` logic is not flawless, it fails when we're using a cache eagerly initialized
    # (e.g. compiled prefill) AND `pixel_values` are not provided (i.e. the image data is provided through other
    # means). Determining prefill in that case requires checking data values, which is not compile-compatible.
    is_first_iteration = (
        is_first_iteration
        if is_first_iteration is not None
        else (past_key_values is None or not past_key_values.is_initialized or pixel_values is not None)
    )
    if mm_token_type_ids is not None and is_first_iteration:
        # We need to pass an additional mask function to account for token type ids, and it needs to be an `or` (to
        # undo the causal masking)

        # First find where a new vision block starts. Vision tokens cannot attend to
        # future vision tokens, but can attend to all prev tokens and to itself bidirectionally
        is_vision = (mm_token_type_ids == 1) | (mm_token_type_ids == 2)
        is_prev_vision = torch.roll(is_vision, shifts=1, dims=-1)
        is_prev_vision[..., 0] = False
        new_vision_starts = is_vision & ~is_prev_vision
        vision_group_ids = torch.cumsum(new_vision_starts.int(), dim=1) - 1
        vision_group_ids = torch.where(is_vision, vision_group_ids, -1)
        sliding_mask_kwargs["or_mask_function"] = token_type_ids_mask_function(
            mm_token_type_ids.to(inputs_embeds.device), vision_group_ids
        )

    return {
        "full_attention": create_causal_mask(**mask_kwargs),
        "sliding_attention": create_sliding_window_causal_mask(**sliding_mask_kwargs),
    }


@auto_docstring(
    custom_intro="""
    The base Gemma 4 model comprising a vision backbone, an audio backbone, and a language model without a
    language modeling head.
    """
)
class Gemma4Model(Gemma4PreTrainedModel):
    # we are filtering the logits/labels so we shouldn't divide the loss based on num_items_in_batch
    accepts_loss_kwargs = False

    def __init__(self, config: Gemma4Config):
        super().__init__(config)
        self.vocab_size = config.text_config.vocab_size

        language_model = AutoModel.from_config(config=config.text_config)
        self.language_model = language_model
        self.vocab_size_per_layer_input = config.text_config.vocab_size_per_layer_input
        self.vision_tower = AutoModel.from_config(config.vision_config) if config.vision_config is not None else None
        self.embed_vision = (
            Gemma4MultimodalEmbedder(config.vision_config, config.text_config)
            if config.vision_config is not None
            else None
        )
        self.audio_tower = AutoModel.from_config(config.audio_config) if config.audio_config is not None else None
        self.embed_audio = (
            Gemma4MultimodalEmbedder(config.audio_config, config.text_config)
            if config.audio_config is not None
            else None
        )
        self.post_init()

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

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

    @can_return_tuple
    @auto_docstring(custom_intro="Projects the last hidden state from the vision model into language model space.")
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_position_ids: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPooling:
        r"""
        image_position_ids (`torch.LongTensor` of shape `(batch_size, max_patches, 2)`, *optional*):
            The patch positions as (x, y) coordinates in the image. Padding patches are indicated by (-1, -1).
        """
        vision_outputs = self.vision_tower(
            pixel_values=pixel_values,
            pixel_position_ids=image_position_ids,
            **kwargs,
        )
        last_hidden_state = vision_outputs.last_hidden_state
        vision_outputs.pooler_output = self.embed_vision(inputs_embeds=last_hidden_state)
        return vision_outputs

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
    ) -> tuple[torch.BoolTensor, torch.BoolTensor, torch.BoolTensor]:
        """
        Obtains mask for multimodal placeholders (replaced by soft tokens) and hard text tokens.

        Masks will be obtained from `mm_token_type_ids`, `input_ids`, or `inputs_embeds` as available and in that
        precedence order. If passing `input_ids` or `inputs_embeds`, the image mask will be derived using
        `config.image_token_id`. Same goes for audio and video masks

        Args:
            input_ids: A tensor containing the hard token IDs from the text tokenizer.
            inputs_embeds: A tensor containing the embeddings for all hard text tokens.

        Returns:
            image_mask, video_mask, audio_mask
        """
        if input_ids is not None:
            special_image_mask = input_ids == self.config.image_token_id
            special_video_mask = input_ids == self.config.video_token_id
            special_audio_mask = input_ids == self.config.audio_token_id
        else:
            special_image_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            ).all(-1)
            special_video_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            ).all(-1)
            special_audio_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.audio_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            ).all(-1)

        return special_image_mask, special_video_mask, special_audio_mask

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        input_features: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        input_features_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        mm_token_type_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        image_position_ids: torch.LongTensor | None = None,
        video_position_ids: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Gemma4ModelOutputWithPast:
        r"""
        input_features_mask (`torch.FloatTensor]` of shape `(num_images, seq_length)`):
            The attention mask for the input audio.
        image_position_ids (`torch.LongTensor` of shape `(batch_size, max_patches, 2)`, *optional*):
            2D patch position coordinates from the image processor, with `(-1, -1)` indicating padding.
            Passed through to the vision encoder for positional embedding computation.
        video_position_ids (`torch.LongTensor` of shape `(num_videos, num_frames, max_patches, 2)`, *optional*):
            2D patch position coordinates from the video processor, with `(-1, -1)` indicating padding.
            Passed through to the vision encoder for positional embedding computation.
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        image_mask, video_mask, audio_mask = self.get_placeholder_mask(input_ids, inputs_embeds)
        multimodal_mask = image_mask | video_mask | audio_mask

        # Replace image id with PAD if the image token if OOV, to avoid index-errors
        llm_input_ids = None
        if inputs_embeds is None:
            llm_input_ids = input_ids.clone()
            llm_input_ids[multimodal_mask] = self.config.text_config.pad_token_id
            inputs_embeds = self.get_input_embeddings()(llm_input_ids)

        if self.config.get_text_config().hidden_size_per_layer_input:
            pad_embedding = self.language_model.embed_tokens.weight[self.config.text_config.pad_token_id, :]
            llm_inputs_embeds = torch.where(multimodal_mask[..., None], pad_embedding.view(1, 1, -1), inputs_embeds)
            per_layer_inputs = self.language_model.get_per_layer_inputs(llm_input_ids, llm_inputs_embeds)
        else:
            per_layer_inputs = None

        # Merge text and images
        if pixel_values is not None:
            image_features = self.get_image_features(pixel_values, image_position_ids, return_dict=True).pooler_output
            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)

            # Confirm the number of soft tokens from the vision tower matches the number of slots in the embeddings.
            n_image_tokens = image_mask.sum()
            image_mask = image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
            torch_compilable_check(
                inputs_embeds[image_mask].numel() == image_features.numel(),
                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features:"
                f" {image_features.shape[0]}",
            )

            inputs_embeds = inputs_embeds.masked_scatter(
                image_mask.to(inputs_embeds.device), image_features.to(inputs_embeds.device)
            )

        if pixel_values_videos is not None:
            video_features = self.get_video_features(
                pixel_values_videos, video_position_ids, return_dict=True
            ).pooler_output
            video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)

            # Confirm the number of soft tokens from the vision tower matches the number of slots in the embeddings.
            n_video_tokens = video_mask.sum()
            video_mask = video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
            torch_compilable_check(
                inputs_embeds[video_mask].numel() == video_features.numel(),
                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features:"
                f" {video_features.shape[0]}",
            )

            inputs_embeds = inputs_embeds.masked_scatter(
                video_mask.to(inputs_embeds.device), video_features.to(inputs_embeds.device)
            )

        # Merge text and audio
        if input_features is not None and input_features_mask is not None:
            audio_output = self.get_audio_features(input_features, input_features_mask, return_dict=True)
            audio_features = audio_output.pooler_output
            audio_mask_from_encoder = audio_output.attention_mask  # True = valid

            # Strip padding tokens: only keep real (non-padding) audio soft tokens.
            # audio_mask_from_encoder is True for valid positions, False for padding tokens.
            # This mirrors the vision encoder's padding stripping (see Gemma4VisionEncoder.forward).
            audio_features = audio_features[audio_mask_from_encoder]

            n_audio_tokens = audio_mask.sum()
            audio_mask = audio_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
            torch_compilable_check(
                inputs_embeds[audio_mask].numel() == audio_features.numel(),
                f"Audio features and audio tokens do not match, tokens: {n_audio_tokens}, features:"
                f" {audio_features.shape[0] * audio_features.shape[1]}",
            )

            inputs_embeds = inputs_embeds.masked_scatter(
                audio_mask.to(inputs_embeds.device), audio_features.to(inputs_embeds.device)
            )

        # It may already have been prepared by, e.g., `generate`
        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.unsqueeze(0)

        if not isinstance(causal_mask_mapping := attention_mask, dict):
            if self.config.get_text_config().use_bidirectional_attention == "vision":
                # Larger Gemma 4 models use Gemma 3's bidirectional attention mask for vision inputs
                causal_mask_mapping = create_causal_mask_mapping(
                    self.config,
                    inputs_embeds,
                    attention_mask,
                    past_key_values,
                    position_ids,
                    mm_token_type_ids,
                    pixel_values,
                    is_training=self.training,
                )
            else:
                # Smaller Gemma models use a conventional casual attention mask
                causal_mask_mapping = create_masks_for_generate(
                    self.config,
                    inputs_embeds,
                    attention_mask,
                    past_key_values,
                    position_ids,
                )

        outputs = self.language_model(
            per_layer_inputs=per_layer_inputs,
            attention_mask=causal_mask_mapping,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            return_dict=True,
            **kwargs,
        )

        return Gemma4ModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
            audio_hidden_states=audio_features if input_features is not None else None,
        )

    @can_return_tuple
    @auto_docstring(custom_intro="Projects the last hidden state from the audio encoder into language model space.")
    def get_audio_features(
        self,
        input_features: torch.Tensor,
        input_features_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Gemma4AudioModelOutput:
        r"""
        input_features (`torch.FloatTensor]` of shape `(num_images, seq_length, num_features)`):
            The tensors corresponding to the input audio.
        input_features_mask (`torch.FloatTensor]` of shape `(num_images, seq_length)`):
            The attention mask for the input audio.
        """
        if self.audio_tower is None:
            raise ValueError(
                "Audio features were requested, but the model was initialized without an audio_config. "
                "Cannot process audio without an audio tower and audio embedder."
            )

        audio_outputs = self.audio_tower(input_features, input_features_mask, return_dict=True, **kwargs)
        audio_outputs.pooler_output = self.embed_audio(inputs_embeds=audio_outputs.last_hidden_state)

        return audio_outputs

    @can_return_tuple
    @auto_docstring(custom_intro="Projects the last hidden state from the vision encoder into language model space.")
    def get_video_features(
        self,
        pixel_values_videos: torch.FloatTensor,
        video_position_ids: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPooling:
        r"""
        video_position_ids (`torch.LongTensor` of shape `(num_videos, num_frames, max_patches, 2)`, *optional*):
            2D patch position coordinates from the video processor, with `(-1, -1)` indicating padding.
            Passed through to the vision encoder for positional embedding computation.
        """
        pixel_values_videos = pixel_values_videos.flatten(0, 1)
        video_position_ids = video_position_ids.flatten(0, 1)
        vision_outputs = self.vision_tower(
            pixel_values=pixel_values_videos,
            pixel_position_ids=video_position_ids,
            **kwargs,
        )
        last_hidden_state = vision_outputs.last_hidden_state
        vision_outputs.pooler_output = self.embed_vision(inputs_embeds=last_hidden_state)
        return vision_outputs


@auto_docstring(
    custom_intro="""
    The base Gemma 4 model comprising a vision backbone, an audio backbone, a language model, and a language modeling
    head.
    """
)
class Gemma4ForConditionalGeneration(Gemma4PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
    base_model_prefix = "model"

    def __init__(self, config: Gemma4Config):
        super().__init__(config)
        self.model = Gemma4Model(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.post_init()

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

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

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_position_ids: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ):
        r"""
        image_position_ids (`torch.LongTensor` of shape `(batch_size, max_patches, 2)`, *optional*):
            2D patch position coordinates from the image processor, with `(-1, -1)` indicating padding.
            Passed through to the vision encoder for positional embedding computation.
        """
        return self.model.get_image_features(pixel_values, image_position_ids, **kwargs)

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        input_features: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        input_features_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        image_position_ids: torch.LongTensor | None = None,
        video_position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        mm_token_type_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Gemma4CausalLMOutputWithPast:
        r"""
        input_features_mask (`torch.FloatTensor]` of shape `(num_images, seq_length)`):
            The attention mask for the input audio.
        image_position_ids (`torch.LongTensor` of shape `(batch_size, max_patches, 2)`, *optional*):
            2D patch position coordinates from the image processor, with `(-1, -1)` indicating padding.
            Passed through to the vision encoder for positional embedding computation.
        video_position_ids (`torch.LongTensor` of shape `(num_videos, num_frames, max_patches, 2)`, *optional*):
            2D patch position coordinates from the video processor, with `(-1, -1)` indicating padding.
            Passed through to the vision encoder for positional embedding computation.
        """
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            input_features=input_features,
            attention_mask=attention_mask,
            input_features_mask=input_features_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            mm_token_type_ids=mm_token_type_ids,
            inputs_embeds=inputs_embeds,
            labels=labels,
            use_cache=use_cache,
            image_position_ids=image_position_ids,
            video_position_ids=video_position_ids,
            return_dict=True,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        if (final_logit_softcapping := self.config.get_text_config().final_logit_softcapping) is not None:
            logits = logits / final_logit_softcapping
            logits = torch.tanh(logits)
            logits = logits * final_logit_softcapping

        loss = None
        if labels is not None:
            # Upcast to float if we need to compute the loss to avoid potential precision issues
            logits = logits.float()
            shift_logits = logits[..., :-1, :]
            shift_labels = labels[..., 1:]
            if attention_mask is not None:
                # we use the input attention mask to shift the logits and labels, because it is 2D.
                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
            else:
                shift_logits = shift_logits.contiguous()
                shift_labels = shift_labels.contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()

            flat_logits = shift_logits.view(-1, self.config.get_text_config().vocab_size)
            flat_labels = shift_labels.view(-1).to(shift_logits.device)
            loss = loss_fct(flat_logits, flat_labels)

        return Gemma4CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
            audio_hidden_states=outputs.audio_hidden_states,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        position_ids=None,
        pixel_values=None,
        pixel_values_videos=None,
        input_features=None,
        attention_mask=None,
        input_features_mask=None,
        token_type_ids=None,
        use_cache=True,
        logits_to_keep=None,
        labels=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- custom `position_ids` and `pixel_values` handling
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            use_cache=use_cache,
            logits_to_keep=logits_to_keep,
            token_type_ids=token_type_ids,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        # If we're in cached decoding stage, multimodal inputs are already cached and can be dropped
        if is_first_iteration or not use_cache:
            model_inputs["pixel_values"] = pixel_values
            model_inputs["pixel_values_videos"] = pixel_values_videos
            model_inputs["input_features"] = input_features
            model_inputs["input_features_mask"] = input_features_mask

        return model_inputs

    @staticmethod
    def create_masks_for_generate(
        config: PreTrainedConfig,
        inputs_embeds: torch.Tensor,
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None,
        position_ids: torch.Tensor | None,
        mm_token_type_ids: torch.Tensor | None = None,
        is_first_iteration: bool | None = False,
        **kwargs,
    ) -> dict:
        if getattr(config.get_text_config(), "use_bidirectional_attention", None) == "vision":
            # Larger Gemma 4 models use Gemma 3's bidirectional attention mask for vision inputs
            return create_causal_mask_mapping(
                config,
                inputs_embeds,
                attention_mask,
                past_key_values,
                position_ids,
                mm_token_type_ids,
                is_first_iteration=is_first_iteration,
                **{k: v for k, v in kwargs.items() if k != "pixel_values"},
            )
        else:
            # Smaller Gemma models use a conventional casual attention mask
            return create_masks_for_generate(
                config, inputs_embeds, attention_mask, past_key_values, position_ids, **kwargs
            )


__all__ = [
    "Gemma4AudioModel",
    "Gemma4ForCausalLM",
    "Gemma4ForConditionalGeneration",
    "Gemma4Model",
    "Gemma4PreTrainedModel",
    "Gemma4TextModel",
    "Gemma4VisionModel",
]