#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/kimi_linear/modular_kimi_linear.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_kimi_linear.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 The HuggingFace Inc. team
#
# 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

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import (
    use_experts_implementation,
    use_kernel_forward_from_hub,
    use_kernel_func_from_hub_with_fallback,
    use_kernelized_func,
)
from ...integrations.accelerate import force_accelerate_hooks
from ...masking_utils import create_causal_mask, create_recurrent_attention_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
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 merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_kimi_linear import KimiLinearConfig


@use_kernel_forward_from_hub("RMSNorm")
class KimiLinearRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        KimiLinearRMSNorm 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}"


# NOTE: the FLA package does not re-cast to `input_dtype` in its implementation, maybe we should do the same
@use_kernel_forward_from_hub("RMSNormGated")
class KimiLinearRMSNormGated(nn.Module):
    def __init__(self, hidden_size, eps=1e-6, **kwargs) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps
        self.activation = "sigmoid"

    def forward(self, hidden_states, gate=None) -> torch.Tensor:
        input_dtype = hidden_states.dtype

        # Strict FP32 norm (do not downcast on the weights)
        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)
        hidden_states = self.weight.to(torch.float32) * hidden_states

        # Apply gating
        hidden_states = hidden_states * ACT2FN[self.activation](gate.to(torch.float32))

        return hidden_states.to(input_dtype)


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

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

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

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

        return final_hidden_states


class KimiLinearMLP(nn.Module):
    def __init__(self, config, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


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

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

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

    return attn_output, attn_weights


class KimiLinearAttention(nn.Module):
    """Multi-headed Latent Attention (MLA) from Deepseek V2 with NoPE, but the part of the keys where RoPE is applied is
    still shared."""

    def __init__(self, config: KimiLinearConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads

        self.q_lora_rank = config.q_lora_rank
        self.qk_rope_head_dim = config.qk_rope_head_dim
        self.kv_lora_rank = config.kv_lora_rank
        self.v_head_dim = config.v_head_dim
        self.qk_nope_head_dim = config.qk_nope_head_dim
        self.qk_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads

        self.is_causal = True

        self.q_proj = (
            None
            if self.q_lora_rank is not None
            else nn.Linear(self.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
        )
        self.q_a_proj = (
            nn.Linear(self.hidden_size, self.q_lora_rank, bias=config.attention_bias)
            if self.q_lora_rank is not None
            else None
        )
        self.q_a_layernorm = KimiLinearRMSNorm(self.q_lora_rank) if self.q_lora_rank is not None else None
        self.q_b_proj = (
            nn.Linear(self.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
            if self.q_lora_rank is not None
            else None
        )

        self.kv_a_proj_with_mqa = nn.Linear(
            self.hidden_size,
            config.kv_lora_rank + config.qk_rope_head_dim,
            bias=config.attention_bias,
        )
        self.kv_a_layernorm = KimiLinearRMSNorm(config.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
        )

        self.o_proj = nn.Linear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=config.attention_bias,
        )
        self.scaling = self.qk_head_dim ** (-0.5)

    def expand_kv(self, kv_nope: torch.Tensor, k_rot: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """Expands the compressed latents into key and value states. Args:
            - kv_nope: key + value without positional encoding, shape [batch_size, 1, seqlen, self.kv_lora_rank]
            - k_rot: shared key with positional encoding, shape [batch_size, 1, seqlen, self.qk_rope_head_dim]
        Returns the key and value states, two tensors of shape [batch, num_heads, seq, (k or v)_head_dim].
        """
        batch_size, _, seq_length, _ = kv_nope.shape
        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)

        kv_nope = self.kv_b_proj(kv_nope).view(key_shape).transpose(1, 2)
        k_nope, value_states = torch.split(kv_nope, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_rot = k_rot.expand(-1, k_nope.shape[1], -1, -1)  # does not affect the underlying storage

        # Concatenate k_nope and k_rot in a new tensor
        key_states = kv_nope.new_empty(*kv_nope.shape[:-1], self.qk_nope_head_dim + self.qk_rope_head_dim)
        key_states[..., : self.qk_nope_head_dim].copy_(k_nope)
        key_states[..., self.qk_nope_head_dim :].copy_(k_rot)
        return key_states, value_states

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
        batch_size, seq_length = hidden_states.shape[:-1]
        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)

        if self.q_lora_rank is None:
            q_states = self.q_proj(hidden_states)
        else:
            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
        query_states = q_states.view(query_shape).transpose(1, 2)

        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
        kv_nope, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        kv_nope = self.kv_a_layernorm(kv_nope)
        # Both latents are viewed as single-head, 4D tensors so all cache layers handle them correctly
        kv_nope = kv_nope.view(batch_size, 1, seq_length, self.kv_lora_rank)
        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)

        # Cache read / write is performed while latent KV is still compressed
        if past_key_values is not None:
            kv_nope, k_rot = past_key_values.update(kv_nope, k_rot, self.layer_idx)

        key_states, value_states = self.expand_kv(kv_nope, k_rot)

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

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class KimiLinearForgetGate(nn.Module):
    """Same as Glm5NextTextForgetGate but with no gate_lower_bound and no A_log reshape."""

    def __init__(self, config: KimiLinearConfig):
        super().__init__()
        self.head_dim = config.linear_head_dim
        self.num_heads = config.linear_num_heads
        self.qkv_dim = self.head_dim * self.num_heads

        self.f_a_proj = nn.Linear(config.hidden_size, self.head_dim, bias=False)
        self.f_b_proj = nn.Linear(self.head_dim, self.qkv_dim, bias=False)
        self.dt_bias = nn.Parameter(torch.empty(self.qkv_dim))
        self.A_log = nn.Parameter(torch.empty(1, 1, self.num_heads, 1))

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_shape = (*hidden_states.shape[:2], -1, self.head_dim)

        forget_gate = self.f_b_proj(self.f_a_proj(hidden_states))
        g = (forget_gate.float() + self.dt_bias.float().view(1, 1, -1)).view(hidden_shape)
        A_log = self.A_log.float()
        decay_rate = torch.exp(A_log)

        # Softplus "log(1 + exp(x))" with uper bound restraint to avoid overflows
        # NOTE: Softplus for larger values (e.g. 20+), Softplus(x) == x
        g_softplus = torch.where(g > 20.0, g, torch.log(1.0 + torch.exp(g)))

        return -decay_rate * g_softplus


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    # NOTE: attention mask is a 2D boolean tensor
    if attention_mask is not None:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


@use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d")
def causal_conv1d_update(
    hidden_states: torch.Tensor,
    conv_state: torch.Tensor,
    weight: nn.Parameter,
    bias: nn.Parameter | None = None,
    activation: str | None = None,
):
    _, hidden_size, seq_len = hidden_states.shape
    state_len = conv_state.shape[-1]

    hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
    conv_state.copy_(hidden_states_new[:, :, -state_len:])
    out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
    out = out[:, :, -seq_len:]
    if activation is not None:
        out = ACT2FN[activation](out)
    return out.to(hidden_states.dtype)


@use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d")
def causal_conv1d_fn(
    hidden_states: torch.Tensor,
    weight: nn.Parameter,
    bias: nn.Parameter | None = None,
    activation: str | None = None,
    **kwargs,
):
    _, hidden_size, seq_len = hidden_states.shape
    padding = weight.shape[-1] - 1

    out = F.conv1d(
        hidden_states.to(weight.dtype),
        weight=weight.unsqueeze(1),
        bias=bias,
        padding=padding,
        groups=hidden_size,
    )[:, :, :seq_len]
    if activation is not None:
        out = ACT2FN[activation](out)
    return out.to(hidden_states.dtype)


def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
    """
    This function is intended to align with the l2norm implementation in the FLA library.

    # NOTE: FLA compares against `F.normalize` but does + eps instead of max(..., eps) leading to a slight differences
    """
    # main difference to qwen's gdn variation: intentionally use sqrt and / to match original triton
    inv_norm = torch.sqrt((x * x).sum(dim=dim, keepdim=True) + eps)
    return x / inv_norm


@use_kernel_func_from_hub_with_fallback("fused_recurrent_kda", "fla")
def recurrent_kimi_delta_attention(
    query,
    key,
    value,
    g,
    beta,
    initial_state,
    output_final_state,
    use_qk_l2norm_in_kernel=False,
    **kwargs,
):
    # calculations happen in float as states are more susceptible to rounding errors
    initial_dtype = query.dtype
    query, key, value, g, beta = [x.to(torch.float32) for x in (query, key, value, g, beta)]

    # important: FLA calculates these in fp32 so we do this after the float casts
    if use_qk_l2norm_in_kernel:
        query = l2norm(query, dim=-1, eps=1e-6)
        key = l2norm(key, dim=-1, eps=1e-6)

    # shapes and other metadata
    batch_size, sequence_length, num_heads, k_head_dim = key.shape
    v_head_dim = value.shape[-1]
    scale = 1 / (query.shape[-1] ** 0.5)
    query = query * scale

    core_attn_out = torch.zeros(
        batch_size, sequence_length, num_heads, v_head_dim, dtype=value.dtype, device=value.device
    )
    last_recurrent_state = (
        torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim, dtype=value.dtype, device=value.device)
        if initial_state is None
        else initial_state.to(value)
    )

    # recurrent iteration
    for i in range(sequence_length):
        q_i = query[:, i]
        k_i = key[:, i]
        v_i = value[:, i]
        g_i = g[:, i][..., None].exp()
        b_i = beta[:, i][..., None]

        last_recurrent_state = last_recurrent_state * g_i
        kv_mem = (last_recurrent_state * k_i[..., None]).sum(dim=-2)
        delta = (v_i - kv_mem) * b_i

        last_recurrent_state = last_recurrent_state + k_i.unsqueeze(-1) * delta.unsqueeze(-2)
        core_attn_out[:, i] = (last_recurrent_state * q_i.unsqueeze(-1)).sum(dim=-2)

    return core_attn_out.to(initial_dtype), last_recurrent_state if output_final_state else None


@use_kernel_func_from_hub_with_fallback("chunk_kda", "fla")
def chunk_kimi_delta_attention(
    query,
    key,
    value,
    g,
    beta,
    chunk_size=64,
    initial_state=None,
    output_final_state=False,
    use_qk_l2norm_in_kernel=False,
    **kwargs,
):
    # calculations happen in float as states are more susceptible to rounding errors
    initial_dtype = query.dtype

    query, key, value, beta, g = [
        x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g)
    ]

    # important: FLA calculates these in fp32 so we do this after the float casts
    if use_qk_l2norm_in_kernel:
        query = l2norm(query, dim=-1, eps=1e-6)
        key = l2norm(key, dim=-1, eps=1e-6)

    # shapes and other metadata
    batch_size, num_heads, sequence_length, k_head_dim = key.shape
    v_head_dim = value.shape[-1]
    scale = 1 / (query.shape[-1] ** 0.5)
    pad_size = (chunk_size - sequence_length % chunk_size) % chunk_size
    total_sequence_length = sequence_length + pad_size

    # prepare all the relevant input
    query = F.pad(query, (0, 0, 0, pad_size)) * scale
    key = F.pad(key, (0, 0, 0, pad_size))
    value = F.pad(value, (0, 0, 0, pad_size))
    g = F.pad(g, (0, 0, 0, pad_size))
    beta = F.pad(beta, (0, pad_size))
    v_beta = value * beta.unsqueeze(-1)
    k_beta = key * beta.unsqueeze(-1)

    # reshape to chunks
    query, key, value, g, k_beta, v_beta = [
        x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1]) for x in (query, key, value, g, k_beta, v_beta)
    ]
    beta = beta.reshape(beta.shape[0], beta.shape[1], -1, chunk_size)

    # Intra chunk
    # Main difference to GDN is the per head application of `g` which was broadcasted across heads instead
    g = g.cumsum(dim=-2)
    mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=0)
    strict_mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1)
    decay_mask = (g.unsqueeze(-2) - g.unsqueeze(-3)).masked_fill(strict_mask[..., None], float("-inf")).exp().float()
    attn = -(k_beta.unsqueeze(-2) * key.unsqueeze(-3) * decay_mask).sum(dim=-1).masked_fill(mask, 0)
    for i in range(1, chunk_size):
        row = attn[..., i, :i].clone()
        sub = attn[..., :i, :i].clone()
        attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)

    attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
    value = attn @ v_beta
    k_cumdecay = attn @ (k_beta * g.exp())

    last_recurrent_state = (
        torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim, dtype=value.dtype, device=value.device)
        if initial_state is None
        else initial_state.to(value)
    )
    core_attn_out = torch.zeros_like(value)

    mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1)
    for i in range(total_sequence_length // chunk_size):
        q_i = query[:, :, i]
        k_i = key[:, :, i]
        v_i = value[:, :, i]
        g_i = g[:, :, i]

        # Inter chunk
        attn_inter = (q_i * g_i.exp()) @ last_recurrent_state
        # Intra chunk
        attn_intra = (q_i.unsqueeze(-2) * k_i.unsqueeze(-3) * decay_mask[:, :, i]).sum(dim=-1).masked_fill(mask, 0)
        # New update rule
        v_prime = k_cumdecay[:, :, i] @ last_recurrent_state
        v_new = v_i - v_prime

        core_attn_out[:, :, i] = attn_inter + attn_intra @ v_new
        last_recurrent_state = (
            last_recurrent_state * g_i[:, :, -1].exp().unsqueeze(-1)
            + (k_i * (g_i[:, :, -1:] - g_i).exp()).transpose(-1, -2) @ v_new
        )

    if not output_final_state:
        last_recurrent_state = None

    core_attn_out = core_attn_out.reshape(core_attn_out.shape[0], core_attn_out.shape[1], -1, core_attn_out.shape[-1])
    core_attn_out = core_attn_out[:, :, :sequence_length]
    core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)

    return core_attn_out, last_recurrent_state


@use_kernelized_func(
    [chunk_kimi_delta_attention, recurrent_kimi_delta_attention, causal_conv1d_fn, causal_conv1d_update]
)
class KimiLinearDeltaAttention(nn.Module):
    """Kimi Linear Attention: this is essentialy the same a gated delta net (GDN) but decay is per-channel instead of
    per-token."""

    def __init__(self, config: KimiLinearConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_heads = config.linear_num_heads
        self.head_dim = config.linear_head_dim
        self.qkv_dim = self.head_dim * self.num_heads

        self.conv_kernel_size = config.linear_conv_kernel_dim
        self.layer_idx = layer_idx
        self.activation = config.hidden_act
        self.layer_norm_epsilon = config.rms_norm_eps

        self.q_proj = nn.Linear(self.hidden_size, self.qkv_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.qkv_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.qkv_dim, bias=False)

        self.conv_dim = self.qkv_dim * 3
        self.conv1d = nn.Conv1d(
            in_channels=self.conv_dim,
            out_channels=self.conv_dim,
            bias=False,
            kernel_size=self.conv_kernel_size,
            groups=self.conv_dim,
            padding=self.conv_kernel_size - 1,
        )
        self.forget_gate = KimiLinearForgetGate(config)
        self.b_proj = nn.Linear(self.hidden_size, self.num_heads, bias=False)

        self.g_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
        self.g_b_proj = nn.Linear(self.head_dim, self.qkv_dim, bias=False)
        self.o_norm = KimiLinearRMSNormGated(self.head_dim, eps=self.layer_norm_epsilon)
        self.o_proj = nn.Linear(self.qkv_dim, self.hidden_size, bias=False)

        self.layer_type = config.layer_types[layer_idx]

    @force_accelerate_hooks("conv1d")
    def forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ):
        # Zero out padding
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)

        # Set up dimensions for reshapes later
        batch_size, seq_len = hidden_states.shape[:2]
        hidden_shape = (batch_size, seq_len, -1, self.head_dim)

        mixed_qkv = torch.cat(
            [
                self.q_proj(hidden_states),
                self.k_proj(hidden_states),
                self.v_proj(hidden_states),
            ],
            dim=-1,
        ).transpose(1, 2)

        # Acts for normal prefill but also for multi-token prefill continue
        use_precomputed_states = cache_params is not None and cache_params.has_previous_state(self.layer_idx)
        if use_precomputed_states:
            conv_state = cache_params.layers[self.layer_idx].conv_states[0]
            recurrent_state = cache_params.layers[self.layer_idx].recurrent_states[0]

        # Single token decode path
        if use_precomputed_states and seq_len == 1:
            mixed_qkv = causal_conv1d_update(
                mixed_qkv,
                conv_state,
                weight=self.conv1d.weight.squeeze(1),
                bias=self.conv1d.bias,
                activation=self.activation,
            )
        # Multi token prefill or simple "full" prefill
        else:
            # Concatenated state for prefill
            if cache_params is not None:
                mixed_qkv = cache_params.update_conv_state(
                    mixed_qkv, self.layer_idx, conv_kernel_size=self.conv_kernel_size
                )

            mixed_qkv = causal_conv1d_fn(
                mixed_qkv,
                weight=self.conv1d.weight.squeeze(1),
                bias=self.conv1d.bias,
                activation=self.activation,
                **kwargs,
            )

            # Cut out any tail
            mixed_qkv = mixed_qkv[:, :, -seq_len:]

        query, key, value = torch.split(
            mixed_qkv.transpose(1, 2),
            [self.qkv_dim] * 3,
            dim=-1,
        )

        query = query.view(hidden_shape)
        key = key.view(hidden_shape)
        value = value.view(hidden_shape)

        # Forget gate and input gate
        g = self.forget_gate(hidden_states)
        beta = torch.sigmoid(self.b_proj(hidden_states))

        # KDA
        if use_precomputed_states and seq_len == 1:
            core_attn_out, last_recurrent_state = recurrent_kimi_delta_attention(
                query,
                key,
                value,
                g=g,
                beta=beta,
                initial_state=recurrent_state,
                output_final_state=cache_params is not None,
                use_qk_l2norm_in_kernel=True,
                **kwargs,
            )
        else:
            core_attn_out, last_recurrent_state = chunk_kimi_delta_attention(
                query,
                key,
                value,
                g=g,
                beta=beta,
                initial_state=recurrent_state if use_precomputed_states else None,
                output_final_state=cache_params is not None,
                use_qk_l2norm_in_kernel=True,
                **kwargs,
            )

        if cache_params is not None:
            cache_params.update_recurrent_state(last_recurrent_state.to(torch.float32), self.layer_idx)

        # Final gated norm and proj
        gate = self.g_b_proj(self.g_a_proj(hidden_states)).view(hidden_shape)
        output = self.o_norm(core_attn_out, gate).reshape(batch_size, seq_len, -1)
        output = self.o_proj(output)

        return output


class KimiLinearTopkRouter(nn.Module):
    def __init__(self, config: KimiLinearConfig):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
        self.routed_scaling_factor = config.routed_scaling_factor
        self.num_group = config.n_group
        self.topk_group = config.topk_group
        self.norm_topk_prob = config.norm_topk_prob
        self.e_score_correction_bias = nn.Buffer(torch.zeros(self.num_experts))

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.hidden_dim)
        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
        scores = router_logits.sigmoid()
        scores_for_choice = scores + self.e_score_correction_bias
        group_scores = (
            scores_for_choice.view(-1, self.num_group, self.num_experts // self.num_group)
            .topk(2, dim=-1)[0]
            .sum(dim=-1)
        )
        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
        group_mask = torch.zeros_like(group_scores)
        group_mask.scatter_(1, group_idx, 1)
        score_mask = (
            group_mask.unsqueeze(-1)
            .expand(-1, self.num_group, self.num_experts // self.num_group)
            .reshape(-1, self.num_experts)
        )
        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), float("-inf"))
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        topk_weights = scores.gather(1, topk_indices)
        if self.norm_topk_prob:
            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
            topk_weights /= denominator
        topk_weights = topk_weights * self.routed_scaling_factor
        return router_logits, topk_weights, topk_indices


class KimiLinearMoE(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config: KimiLinearConfig):
        super().__init__()
        self.config = config
        self.experts = KimiLinearExperts(config)
        self.gate = KimiLinearTopkRouter(config)
        self.shared_experts = KimiLinearMLP(
            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        residuals = hidden_states
        orig_shape = hidden_states.shape
        _, topk_weights, topk_indices = self.gate(hidden_states)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
        hidden_states = hidden_states + self.shared_experts(residuals)
        return hidden_states


class KimiLinearDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: KimiLinearConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = (
            # CODEPATH: TODO: remove this once the mlinter rule is relaxed
            KimiLinearAttention(config, layer_idx)
            if config.layer_types[layer_idx] == "full_attention"
            else KimiLinearDeltaAttention(config, layer_idx)
        )

        self.mlp = KimiLinearMoE(config) if config.mlp_layer_types[layer_idx] == "sparse" else KimiLinearMLP(config)

        self.input_layernorm = KimiLinearRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.post_attention_layernorm = KimiLinearRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.block_type = config.layer_types[layer_idx]

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        # Self attn
        hidden_states = self.input_layernorm(hidden_states)
        if self.block_type == "linear_attention":
            hidden_states = self.self_attn(
                hidden_states=hidden_states,
                cache_params=past_key_values,
                attention_mask=attention_mask,
                **kwargs,
            )
        else:
            hidden_states, _ = self.self_attn(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )
        hidden_states = residual + hidden_states

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


@auto_docstring
class KimiLinearPreTrainedModel(PreTrainedModel):
    config: KimiLinearConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["KimiLinearDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _keys_to_ignore_on_load_unexpected = None
    _can_record_outputs = {
        "router_logits": OutputRecorder(KimiLinearTopkRouter, index=0),
        "hidden_states": KimiLinearDecoderLayer,
        "attentions": KimiLinearAttention,
    }
    _is_stateful = True
    _can_compile_fullgraph = True

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, KimiLinearForgetGate):  # following FLA initialization
            # A_log
            init.copy_(module.A_log, init.uniform_(module.A_log, a=1.0, b=16.0).log())
            # dt_bias
            init.uniform_(module.dt_bias, a=math.log(1e-3), b=math.log(1e-1))
            dt = module.dt_bias.exp().clamp_min(1e-4)
            init.copy_(module.dt_bias, dt + torch.log(-torch.expm1(-dt)))  # (stable) inverse softplus
        elif isinstance(module, KimiLinearExperts):
            init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
            init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, KimiLinearTopkRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.e_score_correction_bias)
        elif isinstance(module, KimiLinearRMSNormGated):
            init.ones_(module.weight)


@auto_docstring
class KimiLinearModel(KimiLinearPreTrainedModel):
    def __init__(self, config: KimiLinearConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id)
        self.layers = nn.ModuleList(
            [KimiLinearDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = KimiLinearRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.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.LongTensor = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device)
            position_ids = (position_ids + past_seen_tokens).unsqueeze(0)

        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # 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),
                "linear_attention": create_recurrent_attention_mask(**mask_kwargs),
            }

        hidden_states = inputs_embeds

        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[self.config.layer_types[i]],
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class KimiLinearForCausalLM(KimiLinearPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    _fsdp_plan = {"lm_head": "keep_full_weight"}

    def __init__(self, config):
        super().__init__(config)
        self.model = KimiLinearModel(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],
    ) -> MoeCausalLMOutputWithPast:
        r"""
        Example:

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

        >>> model = KimiLinearForCausalLM.from_pretrained("meta-kimi_linear/KimiLinear-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-kimi_linear/KimiLinear-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."
        ```"""
        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,
            **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, :])

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


__all__ = ["KimiLinearPreTrainedModel", "KimiLinearModel", "KimiLinearForCausalLM"]