#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/granitemoehybrid/modular_granitemoehybrid.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_granitemoehybrid.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 IBM 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 TypedDict

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_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast, 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.deprecation import deprecate_kwarg
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_granitemoehybrid import GraniteMoeHybridConfig


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)


@use_kernel_forward_from_hub("rotary_pos_emb")
def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

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


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


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

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

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

    return attn_output, attn_weights


@use_kernelized_func(apply_rotary_pos_emb)
class GraniteMoeHybridAttention(nn.Module):
    """Hybrid variant that handles ``position_embeddings is None`` — granitemoe-hybrid configs can
    opt out of RoPE via ``position_embedding_type=None``, in which case the model passes ``None``
    instead of a ``(cos, sin)`` tuple."""

    def __init__(self, config: GraniteMoeHybridConfig, 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.scaling = config.attention_multiplier  # Only diff with llama
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

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

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

        if position_embeddings is not None:
            cos, sin = position_embeddings
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

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

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


# Helper methods for segment sum computation


def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
    """
    Padding x tensor with `pad_size` on the seq_len dim (dim=1)

    Assumes that we only have tensors of either size 4 or 3
    """
    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)

    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)


def reshape_into_chunks(input_tensor, pad_size, chunk_size):
    """
    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
    simultaneously splitting it into chunk sequences.

    Assumes that we only have tensors of either size 4 or 3
    """
    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
    input_tensor = pad_tensor_by_size(input_tensor, pad_size)

    if len(input_tensor.shape) == 3:
        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
    else:
        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
        return input_tensor.reshape(
            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
        )


def segment_sum(input_tensor):
    """
    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
    """
    chunk_size = input_tensor.size(-1)
    # 1. expand input tensor to have an additional dimension and repeat along that dimension
    # [..., chunk_size] -> [..., chunk_size, chunk_size]
    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
    input_tensor = input_tensor.masked_fill(~mask, 0)
    # 3. compute actual cumsum
    tensor_segsum = torch.cumsum(input_tensor, dim=-2)

    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
    return tensor_segsum


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)


@use_kernel_func_from_hub_with_fallback("mamba_split_conv1d_scan_combined", "mamba_ssm")
def mamba2_split_conv1d_scan_combined(
    zxbcdt: torch.Tensor,
    conv1d_weight: torch.Tensor,
    conv1d_bias: torch.Tensor | None,
    dt_bias: torch.Tensor,
    A: torch.Tensor,
    D: torch.Tensor,
    chunk_size: int,
    initial_states: torch.Tensor | None = None,
    dt_limit: tuple[float, float] = (0.0, float("inf")),
    return_final_states: bool = False,
    activation: str = "silu",
    rmsnorm_weight: torch.Tensor | None = None,
    rmsnorm_eps: float = 1e-6,
    outproj_weight: torch.Tensor | None = None,
    outproj_bias: torch.Tensor | None = None,
    headdim: int | None = None,
    ngroups: int = 1,
    norm_before_gate: bool = True,
    **kwargs,
):
    return None


@use_kernel_func_from_hub_with_fallback("selective_state_update", "mamba_ssm")
def mamba2_selective_state_update(
    state: torch.Tensor,
    hidden_states: torch.Tensor,
    dt: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
    D: torch.Tensor | None = None,
    dt_bias: torch.Tensor | None = None,
    dt_softplus: bool = False,
    z: torch.Tensor | None = None,
    **kwargs,
):
    batch_size, num_heads, head_dim = hidden_states.shape
    num_groups = B.shape[1]
    state_size = B.shape[-1]

    if dt_bias is not None:
        dt = dt + dt_bias.to(dt.dtype)
    if dt_softplus:
        dt = F.softplus(dt)
    dt = dt[..., None]

    # Discretize A
    dA = torch.exp(dt.float() * A.float()).to(device=state.device)

    # Discretize B
    B = B.reshape(batch_size, num_groups, 1, state_size)
    B = B.expand(batch_size, num_groups, num_heads // num_groups, state_size).contiguous()
    B = B.reshape(batch_size, num_heads, 1, state_size)
    dB = dt * B

    # Discretize x into dB
    dBx = (dB * hidden_states[..., None]).to(device=state.device)

    # State calculation
    ssm_states = state * dA + dBx
    state.copy_(ssm_states.to(state.dtype))

    # Subsequent output
    C = C.reshape(batch_size, num_groups, 1, state_size)
    C = C.expand(batch_size, num_groups, num_heads // num_groups, state_size).contiguous()
    C = C.reshape(batch_size, num_heads, state_size)

    # Reshape ssm_states to merge the first two dimensions
    ssm_states = ssm_states.to(device=C.device, dtype=C.dtype)
    ssm_states_reshaped = ssm_states.view(batch_size * num_heads, head_dim, state_size)
    C_reshaped = C.view(batch_size * num_heads, state_size, 1)
    out = torch.bmm(ssm_states_reshaped, C_reshaped)
    out = out.view(batch_size, num_heads, head_dim)

    # D skip connection
    if D is not None:
        out = (out + hidden_states * D).to(out.dtype)

    if z is not None:
        out = out * F.silu(z)

    return out.to(hidden_states.dtype)


@use_kernel_func_from_hub_with_fallback("mamba_chunk_scan_combined", "mamba_ssm")
def mamba2_chunk_scan(
    hidden_states: torch.Tensor,
    dt: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
    chunk_size: int,
    D: torch.Tensor | None = None,
    dt_bias: torch.Tensor | None = None,
    initial_states: torch.Tensor | None = None,
    dt_softplus: bool = False,
    dt_limit: tuple[float, float] = (0.0, float("inf")),
    return_final_states: bool = False,
    **kwargs,
):
    batch_size, sequence_length, num_heads, head_dim = hidden_states.shape
    num_groups = B.shape[2]

    if dt_bias is not None:
        dt = dt + dt_bias.to(dt.dtype)
    if dt_softplus:
        dt = F.softplus(dt)
    dt = torch.clamp(dt, min=dt_limit[0], max=dt_limit[1])

    hidden_states = hidden_states.float()
    B = B.float().repeat_interleave(num_heads // num_groups, dim=2, output_size=num_heads)
    C = C.float().repeat_interleave(num_heads // num_groups, dim=2, output_size=num_heads)

    pad_size = (chunk_size - sequence_length % chunk_size) % chunk_size
    D_residual = None
    if D is not None:
        D_residual = D[..., None] * pad_tensor_by_size(hidden_states, pad_size)

    # Discretize x and A
    hidden_states = hidden_states * dt[..., None].float()
    A = A.to(hidden_states.dtype) * dt.float()

    # Rearrange into blocks/chunks
    hidden_states, A, B, C = [reshape_into_chunks(tensor, pad_size, chunk_size) for tensor in (hidden_states, A, B, C)]

    A = A.permute(0, 3, 1, 2)
    A_cumsum = torch.cumsum(A, dim=-1)

    # 1. Compute the output for each intra-chunk (diagonal blocks)
    # This is the analog of a causal mask
    L = torch.exp(segment_sum(A))

    # Contraction of C and B to get G (attention-weights like)
    G = (C[:, :, :, None, :, :] * B[:, :, None, :, :, :]).sum(dim=-1)

    # Compute M, equivalent to applying attention mask to weights
    M = (G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]).sum(dim=-1)

    # Compute Y_diag (apply to values)
    Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)

    # 2. Compute the state for each intra-chunk
    # (right term of low-rank factorization of off-diagonal blocks; B terms)
    decay_states = torch.exp(A_cumsum[:, :, :, -1:] - A_cumsum)
    B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
    states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)

    # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
    # (middle term of factorization of off-diag blocks; A terms)
    previous_states = (
        initial_states[:, None].to(dtype=states.dtype, device=states.device)
        if initial_states is not None
        else torch.zeros_like(states[:, :1])
    )
    states = torch.cat([previous_states, states], dim=1)
    decay_chunk = torch.exp(segment_sum(F.pad(A_cumsum[:, :, :, -1], (1, 0)))).transpose(1, 3)
    new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
    states, final_state = new_states[:, :-1], new_states[:, -1]

    # 4. Compute state -> output conversion per chunk
    # (left term of low-rank factorization of off-diagonal blocks; C terms)
    state_decay_out = torch.exp(A_cumsum)
    C_times_states = C[..., None, :] * states[:, :, None, ...]
    Y_off = C_times_states.sum(-1) * state_decay_out.permute(0, 2, 3, 1)[..., None]

    # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
    output = Y_diag + Y_off
    output = output.reshape(batch_size, -1, num_heads, head_dim)

    if D_residual is not None:
        output = output + D_residual

    # Cutting off padded chunks
    if pad_size > 0:
        output = output[:, :sequence_length]

    if return_final_states:
        return output, final_state

    return output


@use_kernelized_func(
    [
        causal_conv1d_fn,
        causal_conv1d_update,
        mamba2_split_conv1d_scan_combined,
        mamba2_selective_state_update,
        mamba2_chunk_scan,
    ]
)
class GraniteMoeHybridMambaLayer(nn.Module):
    """
    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
    and is why Mamba is called **selective** state spaces)

    The are a few differences between this and Mamba2Mixer:
    - The variable use_precomputed_states is slightly different due to the hybrid cache structure
    - Some extra variables that our layer doesn't need have been removed
    """

    def __init__(self, config: GraniteMoeHybridConfig, layer_idx: int, initialize_mixer_weights: bool = True):
        super().__init__()
        self.use_bias = config.mamba_proj_bias
        self.num_heads = config.mamba_n_heads
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.mamba_d_state
        self.conv_kernel_size = config.mamba_d_conv
        self.intermediate_size = int(config.mamba_expand * self.hidden_size)
        self.layer_idx = layer_idx
        self.use_conv_bias = config.mamba_conv_bias
        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]
        self.layer_norm_epsilon = config.rms_norm_eps
        self.n_groups = config.mamba_n_groups
        self.head_dim = config.mamba_d_head
        self.chunk_size = config.mamba_chunk_size

        self.time_step_limit = config.time_step_limit

        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
        self.conv1d = nn.Conv1d(
            in_channels=self.conv_dim,
            out_channels=self.conv_dim,
            bias=config.mamba_conv_bias,
            kernel_size=self.conv_kernel_size,
            groups=self.conv_dim,
            padding=self.conv_kernel_size - 1,
        )
        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
        self.in_proj = nn.Linear(
            self.hidden_size,
            projection_size,
            bias=self.use_bias,
        )
        # selective projection used to make dt, B and C input dependent

        # time step projection (discretization)
        self.dt_bias = nn.Parameter(torch.empty(self.num_heads))

        # S4D real initialization. These are not discretized!
        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
        self.A_log = nn.Parameter(torch.empty(self.num_heads))
        self.norm = GraniteMoeHybridRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon)
        self.D = nn.Parameter(torch.empty(self.num_heads))
        if initialize_mixer_weights and self.dt_bias.device.type != "meta":
            self.init_granitemoehybrid_weights()
        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias)

        self.layer_type = config.layer_types[layer_idx]

    @torch.no_grad()
    def init_granitemoehybrid_weights(self):
        A = torch.arange(1, self.num_heads + 1, device=self.A_log.device, dtype=torch.float32)
        init.copy_(self.A_log, torch.log(A))
        init.ones_(self.D)
        init.ones_(self.dt_bias)

    @force_accelerate_hooks("conv1d")
    def forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ):
        batch_size, seq_len, _ = hidden_states.shape
        dtype = hidden_states.dtype
        use_precomputed_states = cache_params is not None and cache_params.has_previous_state(self.layer_idx)

        # 1. Gated MLP's linear projection
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        projected_states = self.in_proj(hidden_states)

        A = -torch.exp(self.A_log.float())
        fused_kwargs = kwargs | {"dt_limit": self.time_step_limit}
        if self.training and cache_params is None:
            fused_output = mamba2_split_conv1d_scan_combined(
                projected_states,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.dt_bias,
                A,
                D=self.D,
                chunk_size=self.chunk_size,
                activation=self.activation,
                rmsnorm_weight=self.norm.weight,
                rmsnorm_eps=self.norm.variance_epsilon,
                outproj_weight=self.out_proj.weight,
                outproj_bias=self.out_proj.bias,
                headdim=self.head_dim,
                ngroups=self.n_groups,
                norm_before_gate=False,
                return_final_states=False,
                **fused_kwargs,
            )

            # Only kernels can use this shortcircuit, fallback to normal torch otherwise
            if fused_output is not None:
                return fused_output

        gate, hidden_states_B_C, dt = projected_states.split(
            [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
        )

        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]

        # 2. Convolution sequence transformation
        hidden_states_B_C = hidden_states_B_C.transpose(1, 2)
        if use_precomputed_states and seq_len == 1 and not cache_params.layers[self.layer_idx].record_past:
            hidden_states_B_C = causal_conv1d_update(
                hidden_states_B_C,
                conv_state,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                activation=self.activation,
            )
        else:
            if cache_params is not None:
                hidden_states_B_C = cache_params.update_conv_state(
                    hidden_states_B_C,
                    self.layer_idx,
                    conv_kernel_size=self.conv_kernel_size,
                )

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

            if cache_params is not None:
                hidden_states_B_C = hidden_states_B_C[:, :, -seq_len:]

        # 3. SSM transformation
        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C.transpose(1, 2), attention_mask)
        hidden_states, B, C = torch.split(
            hidden_states_B_C,
            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
            dim=-1,
        )

        # Recurrent form
        if use_precomputed_states and seq_len == 1:
            hidden_states = hidden_states.view(batch_size, self.num_heads, self.head_dim)
            dt = dt.transpose(1, 2).expand(-1, -1, self.head_dim)
            A = A[:, None, None].expand(-1, self.head_dim, self.ssm_state_size)
            B = B.view(batch_size, self.n_groups, self.ssm_state_size)
            C = C.view(batch_size, self.n_groups, self.ssm_state_size)
            D = self.D[:, None].expand(-1, self.head_dim)
            dt_bias = self.dt_bias[:, None].expand(-1, self.head_dim)

            scan_output = mamba2_selective_state_update(
                recurrent_state,
                hidden_states,
                dt,
                A,
                B,
                C,
                D,
                z=None,
                dt_bias=dt_bias,
                dt_softplus=True,
                **kwargs,
            )
            scan_output = scan_output.view(batch_size, 1, -1)

        # Chunk form
        else:
            output_final_state = cache_params is not None
            scan_result = mamba2_chunk_scan(
                hidden_states.view(batch_size, seq_len, self.num_heads, self.head_dim),
                dt,
                A,
                B.view(batch_size, seq_len, self.n_groups, self.ssm_state_size),
                C.view(batch_size, seq_len, self.n_groups, self.ssm_state_size),
                chunk_size=self.chunk_size,
                D=self.D,
                z=None,
                return_final_states=output_final_state,
                dt_bias=self.dt_bias,
                dt_softplus=True,
                initial_states=recurrent_state if use_precomputed_states else None,
                dt_limit=self.time_step_limit,
                **kwargs,
            )

            if output_final_state:
                scan_output, final_state = scan_result
                cache_params.update_recurrent_state(final_state, layer_idx=self.layer_idx)
            else:
                scan_output = scan_result

            scan_output = scan_output.reshape(batch_size, seq_len, -1)

        scan_output = self.norm(scan_output, gate)

        # 4. Final linear projection
        contextualized_states = self.out_proj(scan_output.to(dtype))
        return contextualized_states


class GraniteMoeHybridRMSNormGated(torch.nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states, gate=None):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)

        if gate is not None:
            hidden_states = hidden_states * nn.functional.silu(gate.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)


class GraniteMoeHybridMLP(nn.Module):
    """
    MLP layer for shared experts

    Args:
        config:
            Configuration object with model hyperparameters.
    """

    def __init__(self, config: GraniteMoeHybridConfig):
        super().__init__()

        self.input_size = config.hidden_size
        self.hidden_size = config.shared_intermediate_size
        self.activation = ACT2FN[config.hidden_act]
        self.input_linear = nn.Linear(self.input_size, self.hidden_size * 2, bias=False)
        self.output_linear = nn.Linear(self.hidden_size, self.input_size, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.input_linear(hidden_states)
        chunked_hidden_states = hidden_states.chunk(2, dim=-1)
        hidden_states = self.activation(chunked_hidden_states[0]) * chunked_hidden_states[1]
        hidden_states = self.output_linear(hidden_states)
        return hidden_states


class GraniteMoeHybridRotaryEmbedding(nn.Module):
    @deprecate_kwarg("device", version="5.18")
    def __init__(self, config: GraniteMoeHybridConfig, 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.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)

    @staticmethod
    @deprecate_kwarg("device", version="5.18")
    def compute_default_rope_parameters(
        config: GraniteMoeHybridConfig, device=None, **kwargs
    ) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

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

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


class GraniteMoeHybridTopKRouter(nn.Module):
    """Top-k gating that returns the routing decisions without grouping tokens by expert.

    Returns ``(top_k_index, top_k_weights, router_logits)``; the grouping/scattering used to live
    here (via ``expert_size.tolist()``, which broke fullgraph compile) and now happens inside the
    experts forward via ``use_experts_implementation`` so the default ``grouped_mm`` / ``batched_mm``
    paths can compile cleanly.
    """

    def __init__(self, config: GraniteMoeHybridConfig):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.top_k = config.num_experts_per_tok
        self.weight = nn.Parameter(torch.empty(self.num_experts, config.hidden_size))

    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        router_logits = F.linear(hidden_states, self.weight).float()  # (num_tokens, num_experts)
        top_k_logits, top_k_index = router_logits.topk(self.top_k, dim=-1)  # (num_tokens, top_k)
        top_k_weights = torch.softmax(top_k_logits, dim=-1).type_as(hidden_states)  # (num_tokens, top_k)
        return top_k_index, top_k_weights, router_logits


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

    def __init__(self, config: GraniteMoeHybridConfig):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.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 GraniteMoeHybridMoE(nn.Module):
    """Sparsely-gated mixture-of-experts block: router decides, experts compute."""

    def __init__(self, config: GraniteMoeHybridConfig):
        super().__init__()
        self.input_size = config.hidden_size
        self.router = GraniteMoeHybridTopKRouter(config)
        self.experts = GraniteMoeHybridExperts(config)

    def forward(self, layer_input: torch.Tensor) -> torch.Tensor:
        bsz, length, emb_size = layer_input.size()
        hidden_states = layer_input.reshape(-1, emb_size)
        top_k_index, top_k_weights, _ = self.router(hidden_states)
        layer_output = self.experts(hidden_states, top_k_index, top_k_weights)
        return layer_output.view(bsz, length, self.input_size)


class GraniteFlashAttentionKwargs(TypedDict, total=False):
    """
    Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage.
    Use cases include padding-free training and fewer `torch.compile` graph breaks.

    cu_seq_lens_q (`torch.LongTensor`):
        Gets cumulative sequence length for query state.
    cu_seq_lens_k (`torch.LongTensor`):
        Gets cumulative sequence length for key state.
    max_length_q (`int`):
        Maximum sequence length for query state.
    max_length_k (`int`):
        Maximum sequence length for key state.
    seq_idx (`torch.IntTensor):
        Index of each packed sequence.
    """

    cu_seq_lens_q: torch.LongTensor
    cu_seq_lens_k: torch.LongTensor
    max_length_q: int
    max_length_k: int
    seq_idx: torch.IntTensor


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


class GraniteMoeHybridDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: GraniteMoeHybridConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        # Either attention or mamba will be initialized, depending on the layer type.
        self.self_attn = None
        self.input_layernorm = GraniteMoeHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = GraniteMoeHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        # Allow non-MoE (dense)
        self.block_sparse_moe = GraniteMoeHybridMoE(config) if config.num_local_experts > 0 else None
        self.residual_multiplier = config.residual_multiplier  # Only diff with mixtral!
        self.shared_mlp = GraniteMoeHybridMLP(config)
        self.mamba = None

        if config.layers_block_type[layer_idx] == "linear_attention":
            self.mamba = GraniteMoeHybridMambaLayer(config, layer_idx)
        else:
            self.self_attn = GraniteMoeHybridAttention(config, layer_idx)
        self.block_type = config.layers_block_type[layer_idx]

        # Accept 0 experts: skip MoE if num_local_experts == 0
        self.has_experts = getattr(config, "num_local_experts", 0) > 0

    @auto_docstring
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[GraniteFlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        if self.mamba is not None:
            hidden_states = self.mamba(
                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,
                past_key_values=past_key_values,
                use_cache=use_cache,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = residual + hidden_states * self.residual_multiplier
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)

        if self.has_experts:
            moe_hidden_states = self.block_sparse_moe(hidden_states)
            hidden_states = moe_hidden_states + self.shared_mlp(hidden_states)
        else:
            hidden_states = self.shared_mlp(hidden_states)

        hidden_states = residual + hidden_states * self.residual_multiplier
        return hidden_states


@auto_docstring
class GraniteMoeHybridPreTrainedModel(PreTrainedModel):
    config: GraniteMoeHybridConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["GraniteMoeHybridDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": GraniteMoeHybridDecoderLayer,
        "attentions": GraniteMoeHybridAttention,
    }
    _is_stateful = True

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, GraniteMoeHybridExperts):
            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, GraniteMoeHybridTopKRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
        if isinstance(module, GraniteMoeHybridMambaLayer):
            init.ones_(module.dt_bias)
            init.copy_(module.A_log, torch.log(torch.arange(1, module.num_heads + 1)))
            init.ones_(module.D)
        elif isinstance(module, GraniteMoeHybridRMSNormGated):
            init.ones_(module.weight)


@auto_docstring
class GraniteMoeHybridModel(GraniteMoeHybridPreTrainedModel):
    def __init__(self, config: GraniteMoeHybridConfig):
        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(
            [GraniteMoeHybridDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = GraniteMoeHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = GraniteMoeHybridRotaryEmbedding(config) if config.position_embedding_type == "rope" else None
        self.gradient_checkpointing = False
        self.embedding_multiplier = config.embedding_multiplier

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

    @auto_docstring
    @merge_with_config_defaults
    @capture_outputs
    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[GraniteFlashAttentionKwargs],
    ) -> tuple | BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

        inputs_embeds = inputs_embeds * self.embedding_multiplier

        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):
            # Prepare mask arguments
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
                "linear_attention": create_recurrent_attention_mask(**mask_kwargs),
            }

        # embed positions
        hidden_states = inputs_embeds
        position_embeddings = None
        if self.rotary_emb is not None:
            position_embeddings = self.rotary_emb(hidden_states, position_ids)

        for i, decoder_layer in enumerate(self.layers):
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[self.config.layers_block_type[i]],
                past_key_values=past_key_values,
                use_cache=use_cache,
                position_embeddings=position_embeddings,
                **kwargs,
            )
        hidden_states = self.norm(hidden_states)

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


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

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

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

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

    if isinstance(gate_logits, tuple):
        compute_device = gate_logits[0].device
        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)

    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)

    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)

    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)

    if attention_mask is None:
        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.mean(routing_weights, dim=0)
    else:
        batch_size, sequence_length = attention_mask.shape
        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)

        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
        expert_attention_mask = (
            attention_mask[None, :, :, None, None]
            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
            .reshape(-1, top_k, num_experts)
            .to(compute_device)
        )

        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
            expert_attention_mask, dim=0
        )

        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
        router_per_expert_attention_mask = (
            attention_mask[None, :, :, None]
            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
            .reshape(-1, num_experts)
            .to(compute_device)
        )

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
            router_per_expert_attention_mask, dim=0
        )

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


@auto_docstring
class GraniteMoeHybridForCausalLM(GraniteMoeHybridPreTrainedModel, 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"])}
    _fsdp_plan = {"lm_head": "keep_full_weight"}

    def __init__(self, config: GraniteMoeHybridConfig):
        super().__init__(config)
        self.model = GraniteMoeHybridModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.router_aux_loss_coef = config.router_aux_loss_coef
        self.num_experts = config.num_local_experts
        self.num_experts_per_tok = config.num_experts_per_tok
        self.logits_scaling = config.logits_scaling

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

    @auto_docstring
    @can_return_tuple
    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,
        output_router_logits: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,
    ) -> tuple | MoeCausalLMOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

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

        >>> model = GraniteMoeHybridForCausalLM.from_pretrained("ibm-granite/granite-4.0-h-tiny")
        >>> tokenizer = AutoTokenizer.from_pretrained("ibm-granite/granite-4.0-h-tiny")

        >>> 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
        )
        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        # Only compute necessary logits
        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, :])
        logits = logits / self.config.logits_scaling

        loss = None
        if labels is not None:
            # Flatten the tokens
            loss = self.loss_function(
                logits,
                labels,
                vocab_size=self.config.vocab_size,
                **kwargs,
            )

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


__all__ = ["GraniteMoeHybridForCausalLM", "GraniteMoeHybridModel", "GraniteMoeHybridPreTrainedModel"]