#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/zaya/modular_zaya.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_zaya.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 Zyphra and the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


from collections.abc import Callable
from typing import Any, Optional

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

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import use_experts_implementation, use_kernel_forward_from_hub
from ...masking_utils import create_causal_mask, create_recurrent_attention_mask, create_sliding_window_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_zaya import ZayaConfig


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

    def __init__(self, config: ZayaConfig):
        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 = list(set(config.layer_types))
        self.rope_type = {}
        for layer_type in self.layer_types:
            rope_params = self.config.rope_parameters[layer_type]
            if rope_params is None:
                continue

            self.rope_type[layer_type] = rope_params["rope_type"]
            rope_init_fn: Callable = self.compute_default_rope_parameters
            if self.rope_type[layer_type] != "default":
                rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
            curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, layer_type=layer_type)
            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: ZayaConfig | 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).
        """
        base = config.rope_parameters[layer_type]["rope_theta"]
        # key difference to gemma3: partial rope
        partial_rotary_factor = config.rope_parameters[layer_type].get("partial_rotary_factor", 1.0)
        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        dim = int(head_dim * partial_rotary_factor)

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

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

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


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)


class ZayaCCAProjection(nn.Module):
    """
    Projects hidden states into attention q/k/v states with ZAYA's Compressed Convolutional Attention (CCA) path.
    See https://huggingface.co/papers/2510.04476.

    This follows the usual q/k/v projection flow, with three ZAYA-specific changes: q/k are mixed by a causal 1D
    convolution, q/k keep residual projection paths, and v uses a delayed recurrent state.
    """

    def __init__(self, config: ZayaConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx

        self.hidden_size = config.hidden_size

        self.depthwise_kernel_size = config.cca_time0
        self.grouped_kernel_size = config.cca_time1
        self.conv_kernel_size = (self.depthwise_kernel_size - 1) + (self.grouped_kernel_size - 1)

        self.num_key_value_heads = config.num_key_value_heads
        self.num_attention_heads = config.num_attention_heads
        self.head_dim = config.head_dim
        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads

        query_hidden_size = self.num_attention_heads * self.head_dim
        key_value_hidden_size = self.num_key_value_heads * self.head_dim

        self.q_proj = nn.Linear(self.hidden_size, query_hidden_size, bias=self.config.attention_bias)
        self.k_proj = nn.Linear(self.hidden_size, key_value_hidden_size, bias=self.config.attention_bias)
        self.v_proj_current = nn.Linear(self.hidden_size, key_value_hidden_size // 2, bias=self.config.attention_bias)
        self.v_proj_delayed = nn.Linear(self.hidden_size, key_value_hidden_size // 2, bias=self.config.attention_bias)

        conv_channels = key_value_hidden_size + query_hidden_size
        self.conv_qk_depthwise = nn.Conv1d(
            in_channels=conv_channels,
            out_channels=conv_channels,
            kernel_size=self.depthwise_kernel_size,
            groups=conv_channels,
            padding=0,
            stride=1,
        )
        self.conv_qk_grouped = nn.Conv1d(
            in_channels=conv_channels,
            out_channels=conv_channels,
            kernel_size=self.grouped_kernel_size,
            groups=(self.num_key_value_heads + self.num_attention_heads),
            padding=0,
            stride=1,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Cache | None,
        conv_mask: torch.Tensor | None = None,
    ):
        if conv_mask is not None:
            hidden_states = hidden_states * conv_mask[:, :, None].to(hidden_states.dtype)

        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        projected_queries = self.q_proj(hidden_states)
        projected_keys = self.k_proj(hidden_states)
        qk_states = torch.cat([projected_queries, projected_keys], dim=-1)

        query_residual = projected_queries.view(*hidden_shape)
        key_residual = projected_keys.view(*hidden_shape).transpose(1, 2)
        key_residual = repeat_kv(key_residual, self.num_key_value_groups).transpose(1, 2)
        query_residual = (query_residual + key_residual) * 0.5
        key_residual = query_residual.view(*input_shape, -1, self.num_key_value_groups, self.head_dim).mean(dim=-2)

        qk_states = qk_states.transpose(1, 2)
        use_precomputed_states = past_key_values is not None and past_key_values.has_previous_state(self.layer_idx)
        if use_precomputed_states:
            cached_qk_states = past_key_values.layers[self.layer_idx].conv_states[0]
            qk_states = torch.cat([cached_qk_states, qk_states], dim=-1)
        else:
            qk_states = F.pad(qk_states, (self.conv_kernel_size, 0))

        if past_key_values is not None:
            new_conv_state = qk_states[..., -self.conv_kernel_size :]
            new_conv_state = F.pad(new_conv_state, (self.conv_kernel_size - new_conv_state.shape[-1], 0))
            past_key_values.update_conv_state(new_conv_state, self.layer_idx)

        qk_states = self.conv_qk_depthwise(qk_states)
        qk_states = self.conv_qk_grouped(qk_states).transpose(1, 2)

        query_hidden_size = query_residual.shape[-2] * query_residual.shape[-1]
        query = qk_states[..., :query_hidden_size].view(*hidden_shape) + query_residual
        key = qk_states[..., query_hidden_size:].view(*hidden_shape) + key_residual

        # The value path carries half of each value head from the current token and half from the previous token.
        # During cached decoding, `recurrent_v_state` is the previous token's delayed projection.
        value_current = self.v_proj_current(hidden_states)
        delayed_v_state = self.v_proj_delayed(hidden_states)
        if use_precomputed_states:
            recurrent_v_state = past_key_values.layers[self.layer_idx].recurrent_states[0].unsqueeze(1)
        else:
            recurrent_v_state = self.v_proj_delayed(hidden_states.new_zeros(input_shape[0], 1, self.hidden_size))
        value_delayed = torch.cat([recurrent_v_state, delayed_v_state[:, :-1]], dim=1)

        if past_key_values is not None:
            past_key_values.update_recurrent_state(delayed_v_state[:, -1, :], self.layer_idx)

        value = torch.cat([value_current, value_delayed], dim=-1).view(*hidden_shape)

        return query, key, value


class ZayaQKNorm(nn.Module):
    """
    L2-normalizes q/k states to sqrt(head_dim) and applies ZAYA's learned per-KV-head key scale.
    """

    def __init__(self, config: ZayaConfig):
        super().__init__()
        self.head_dim_scale = config.head_dim**0.5
        self.temp = nn.Parameter(torch.zeros(config.num_key_value_heads))

    def forward(self, query_states: torch.Tensor, key_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        norm_eps = torch.finfo(query_states.dtype).eps
        query_states = query_states * (
            self.head_dim_scale / query_states.norm(p=2, dim=-1, keepdim=True).clamp_min(norm_eps)
        )
        key_states = key_states * (
            self.head_dim_scale / key_states.norm(p=2, dim=-1, keepdim=True).clamp_min(norm_eps)
        )
        key_states = key_states * self.temp[None, None, :, None]
        return query_states, key_states


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 eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

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

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

    return attn_output, attn_weights


# Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Removes the interleaving of cos and sin from GLM

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

    # Keep half or full tensor for later concatenation
    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    # Apply rotary embeddings on the first half or full tensor
    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

    # Concatenate back to full shape
    q_embed = torch.cat([q_embed, q_pass], dim=-1)
    k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


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

    def __init__(self, config: ZayaConfig, layer_idx: int):
        super().__init__()
        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.num_key_value_heads = config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.qkv_proj = ZayaCCAProjection(
            config=self.config,
            layer_idx=layer_idx,
        )
        self.layer_type = config.layer_types[layer_idx]
        self.sliding_window = config.sliding_window if self.layer_type == "hybrid_sliding" else None
        self.hidden_size = config.hidden_size
        self.num_attention_heads = config.num_attention_heads
        self.qk_norm = ZayaQKNorm(config)

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

        mask_mapping = attention_mask or {}
        causal_mask = mask_mapping.get("causal")
        conv_mask = mask_mapping.get("conv")

        # ZAYA replaces the usual independent q/k/v projections with CCA projection followed by special q/k normalization.
        query_states, key_states, value_states = self.qkv_proj(hidden_states, past_key_values, conv_mask)
        query_states, key_states = self.qk_norm(query_states, key_states)

        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

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

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

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            causal_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            sliding_window=self.sliding_window,
            **kwargs,
        )

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

        return attn_output, attn_weights


class ZayaDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: ZayaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = ZayaAttention(config=config, layer_idx=layer_idx)
        self.mlp = ZayaSparseMoeBlock(config, layer_idx)
        self.input_layernorm = ZayaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = ZayaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_residual_scale = ZayaResidualScaling(config.hidden_size)
        self.post_mlp_residual_scale = ZayaResidualScaling(config.hidden_size)

    def forward(
        self,
        hidden_states: torch.Tensor,
        prev_router_hidden_states: torch.Tensor | None = None,
        attention_mask: dict[str, Any] | None = None,
        past_key_values: Cache | None = None,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        residual = hidden_states
        # Match upstream's residual_in_fp32 path by keeping the residual stream in fp32 and avoiding extra
        # fp32->bf16 round trips in the residual module.
        hidden_states = self.input_layernorm(residual.to(dtype=self.input_layernorm.weight.dtype))

        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_embeddings=position_embeddings,
            **kwargs,
        )

        residual = self.post_attention_residual_scale(hidden_states, residual)
        hidden_states = self.post_attention_layernorm(residual.to(dtype=self.post_attention_layernorm.weight.dtype))

        hidden_states, prev_router_hidden_states = self.mlp(
            hidden_states,
            prev_router_hidden_states,
        )

        hidden_states = self.post_mlp_residual_scale(hidden_states, residual)

        return hidden_states, prev_router_hidden_states


class ZayaResidualScaling(nn.Module):
    def __init__(self, hidden_size: int):
        super().__init__()
        self.hidden_states_scale = nn.Parameter(torch.ones(hidden_size))
        self.hidden_states_bias = nn.Parameter(torch.zeros(hidden_size))
        self.residual_scale = nn.Parameter(torch.ones(hidden_size))
        self.residual_bias = nn.Parameter(torch.zeros(hidden_size))

    def forward(self, hidden_states: torch.Tensor, residual: torch.Tensor):
        # Keep the residual stream in fp32 to match the original ZAYA `residual_in_fp32` path.
        hidden_states = (hidden_states + self.hidden_states_bias) * self.hidden_states_scale
        residual = (residual + self.residual_bias) * self.residual_scale
        return hidden_states + residual


class ZayaRouterMLP(nn.Module):
    def __init__(self, hidden_size: int, num_experts: int, rms_norm_eps: float):
        super().__init__()
        self.norm = ZayaRMSNorm(hidden_size, eps=rms_norm_eps)
        self.fc1 = nn.Linear(hidden_size, hidden_size, bias=True)
        self.fc2 = nn.Linear(hidden_size, hidden_size, bias=True)
        self.out_proj = nn.Linear(hidden_size, num_experts, bias=False)
        self.act_fn = nn.GELU()

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.norm(hidden_states)
        hidden_states = self.act_fn(self.fc1(hidden_states))
        hidden_states = self.act_fn(self.fc2(hidden_states))
        return self.out_proj(hidden_states)


class ZayaRouter(nn.Module):
    def __init__(
        self,
        config,
        layer_idx: int,
    ) -> None:
        super().__init__()

        self.config = config
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx

        self.num_experts = config.num_experts
        # Zaya1 has a skip expert w/o actual compute
        self.num_router_classes = config.num_experts + 1
        self.top_k = config.num_experts_per_tok
        self.router_hidden_size = config.router_hidden_size

        self.down_proj = nn.Linear(self.hidden_size, self.router_hidden_size, bias=True)

        self.use_eda = self.layer_idx != 0
        if self.use_eda:
            self.router_states_scale = nn.Parameter(torch.ones(self.router_hidden_size))

        self.router_mlp = ZayaRouterMLP(self.router_hidden_size, self.num_router_classes, config.rms_norm_eps)

        self.register_buffer("balancing_biases", torch.zeros(self.num_router_classes, dtype=torch.float32))
        self.balancing_biases[-1] = -1.0

    def forward(
        self,
        hidden_states: torch.Tensor,
        router_states: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        final_shape = (-1, self.top_k)
        seq_length = hidden_states.shape[1]

        router_hidden_states = self.down_proj(hidden_states)

        if self.use_eda and router_states is not None:
            router_hidden_states = router_hidden_states + router_states * self.router_states_scale

        router_hidden_states_next = router_hidden_states[:, -seq_length:].clone()
        router_logits = self.router_mlp(router_hidden_states)
        router_probs = torch.softmax(router_logits, dim=-1)

        biased_router_probs = router_probs.detach().to(torch.float32) + self.balancing_biases
        _, router_indices = torch.topk(biased_router_probs, self.top_k, dim=-1)
        router_probs = torch.gather(router_probs, dim=2, index=router_indices)

        # If the router selects the extra skip expert, mask it before `ZayaExperts` builds its one-hot expert mask.
        skip_expert = router_indices == self.config.num_experts
        router_probs = router_probs.masked_fill(skip_expert, 0)
        router_indices = router_indices.masked_fill(skip_expert, 0)

        return (
            router_logits.reshape(-1, self.num_router_classes),
            router_probs.reshape(final_shape),
            router_indices.reshape(final_shape),
            router_hidden_states_next,
        )


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

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        top_k_index: torch.Tensor,
        top_k_weights: torch.Tensor,
    ) -> torch.Tensor:
        final_hidden_states = torch.zeros_like(hidden_states)
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
            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 ZayaSparseMoeBlock(nn.Module):
    def __init__(self, config, layer_idx: int):
        super().__init__()
        self.gate = ZayaRouter(config, layer_idx)
        self.experts = ZayaExperts(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        prev_router_hidden_states: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        # ZAYA carries router hidden states across decoder layers; the next layer consumes this state in its router.
        _, router_probs, router_indices, prev_router_hidden_states = self.gate(
            hidden_states, router_states=prev_router_hidden_states
        )

        batch_size, seq_length, emb_dim = hidden_states.shape
        hidden_states_flat = hidden_states.view(batch_size * seq_length, emb_dim)
        expert_output = self.experts(hidden_states_flat, router_indices, router_probs)
        expert_output = expert_output.view(batch_size, seq_length, emb_dim)

        return expert_output, prev_router_hidden_states


@auto_docstring
class ZayaPreTrainedModel(PreTrainedModel):
    config: ZayaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["ZayaDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    # ZAYA generation uses the native hybrid dynamic cache, which is not a compileable cache.
    _can_compile_fullgraph = False
    _supports_attention_backend = True
    _can_record_outputs = {
        "router_logits": OutputRecorder(ZayaRouter, index=0),
        "hidden_states": ZayaDecoderLayer,
        "attentions": ZayaAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, ZayaResidualScaling):
            init.ones_(module.hidden_states_scale)
            init.zeros_(module.hidden_states_bias)
            init.ones_(module.residual_scale)
            init.zeros_(module.residual_bias)
        elif isinstance(module, ZayaModel):
            init.ones_(module.input_hidden_states_scale)
            init.zeros_(module.input_hidden_states_bias)
        elif isinstance(module, ZayaQKNorm):
            init.zeros_(module.temp)
        elif isinstance(module, ZayaRouter):
            if module.use_eda:
                init.ones_(module.router_states_scale)
            init.zeros_(module.balancing_biases)
            module.balancing_biases[-1] = -1.0  # trf-ignore: TRF012
        elif isinstance(module, ZayaExperts):
            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, ZayaRotaryEmbedding):
            for layer_type in module.layer_types:
                rope_init_fn = module.compute_default_rope_parameters
                if module.rope_type[layer_type] != "default":
                    rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
                curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
                getattr(module, f"{layer_type}_inv_freq").copy_(curr_inv_freq)
                getattr(module, f"{layer_type}_original_inv_freq").copy_(curr_inv_freq)


@auto_docstring
class ZayaModel(ZayaPreTrainedModel):
    def __init__(self, config: ZayaConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [ZayaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = ZayaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = ZayaRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.input_hidden_states_scale = nn.Parameter(torch.ones(config.hidden_size))
        self.input_hidden_states_bias = nn.Parameter(torch.zeros(config.hidden_size))

        # 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,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

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

        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):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            mask_creation_functions = {
                "hybrid": lambda: create_causal_mask(**mask_kwargs),
                "hybrid_sliding": lambda: create_sliding_window_causal_mask(**mask_kwargs),
            }
            causal_mask_mapping = {
                layer_type: mask_creation_functions[layer_type]() for layer_type in set(self.config.layer_types)
            }
            causal_mask_mapping["conv"] = create_recurrent_attention_mask(**mask_kwargs)

        hidden_states = inputs_embeds

        position_embeddings = {
            layer_type: self.rotary_emb(hidden_states, position_ids, layer_type)
            for layer_type in set(self.config.layer_types)
        }

        # Keep the residual stream in fp32 to match the original ZAYA `residual_in_fp32` path.
        hidden_states = ((hidden_states + self.input_hidden_states_bias) * self.input_hidden_states_scale).to(
            torch.float32
        )

        prev_router_hidden_states = None

        for idx, decoder_layer in enumerate(self.layers):
            layer_type = self.config.layer_types[idx]
            # Attention uses the prepared causal mask, while CCA projection still needs the raw 2D mask to zero padding
            # tokens before convolution.
            hidden_states, prev_router_hidden_states = decoder_layer(
                hidden_states,
                prev_router_hidden_states,
                attention_mask={
                    "causal": causal_mask_mapping[layer_type],
                    "conv": causal_mask_mapping.get("conv"),
                },
                past_key_values=past_key_values,
                position_embeddings=position_embeddings[layer_type],
                **kwargs,
            )

        hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))

        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )


@auto_docstring(checkpoint="Zyphra/ZAYA1-8B")
class ZayaForCausalLM(ZayaPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _is_stateful = True

    def __init__(self, config, **kwargs):
        super().__init__(config)
        self.model = ZayaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=self.config.lm_head_bias)
        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,
        output_router_logits: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeCausalLMOutputWithPast:
        r"""
        Example:

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

        >>> model = ZayaForCausalLM.from_pretrained("meta-zaya/Zaya-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-zaya/Zaya-2-7b-hf")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> 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]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )

        outputs: MoeModelOutputWithPast = 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,
            output_router_logits=output_router_logits,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        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, :])

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

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

    @staticmethod
    def create_masks_for_generate(config, inputs_embeds, attention_mask, past_key_values, position_ids=None, **_):
        # ZAYA decoder layers are hybrid: attention consumes the causal/sliding mask, while CCA uses a 2D padding
        # mask to zero tokens before the convolutional projection.
        text_config = config.get_text_config()
        mask_kwargs = {
            "config": text_config,
            "inputs_embeds": inputs_embeds,
            "attention_mask": attention_mask,
            "past_key_values": past_key_values,
            "position_ids": position_ids,
        }
        mask_creation_functions = {
            "hybrid": lambda: create_causal_mask(**mask_kwargs),
            "hybrid_sliding": lambda: create_sliding_window_causal_mask(**mask_kwargs),
        }
        mask_mapping = {
            layer_type: mask_creation_functions[layer_type]() for layer_type in set(text_config.layer_types)
        }
        mask_mapping["conv"] = create_recurrent_attention_mask(**mask_kwargs)
        return mask_mapping


__all__ = ["ZayaPreTrainedModel", "ZayaModel", "ZayaForCausalLM"]