team-10/venv/Lib/site-packages/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           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.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# 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 typing import Any, Callable, Optional, TypedDict, Union

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

from transformers.activations import ACT2FN

from ...cache_utils import Cache, DynamicCache, DynamicLayer
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import AttentionMaskConverter
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 auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
from ...utils.import_utils import is_causal_conv1d_available, is_mamba_2_ssm_available
from .configuration_granitemoehybrid import GraniteMoeHybridConfig


if is_mamba_2_ssm_available():
    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
    from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
else:
    selective_state_update = None

if is_causal_conv1d_available():
    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
else:
    causal_conv1d_update, causal_conv1d_fn = None, None


if is_torch_flex_attn_available():
    from torch.nn.attention.flex_attention import BlockMask

    from ...integrations.flex_attention import make_flex_block_causal_mask


logger = logging.get_logger(__name__)


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        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: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    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:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    # upcast attention to fp32
    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


# copied from transformers.models.granite.modeling_granite.GraniteAttention with Granite->GraniteMoeHybrid
# no longer copied after attention refactors
class GraniteMoeHybridAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: GraniteMoeHybridConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.is_causal = True

        self.scaling = config.attention_multiplier

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # None or rope embeddings
        **kwargs,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

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

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

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

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        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.view(bsz, q_len, -1)
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights


class HybridMambaAttentionDynamicCache(Cache):
    """
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
    (which has a constant shape regardless of seq_len).

    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
    """

    key_cache = None
    value_cache = None
    is_compileable = False

    def __init__(self, config: GraniteMoeHybridConfig, batch_size, dtype=torch.float16, device=None):
        super().__init__(layer_classes=DynamicLayer)
        self.layers_block_type = config.layers_block_type
        self.has_previous_state = False  # only used by mamba
        conv_kernel_size = config.mamba_d_conv
        ssm_state_size = config.mamba_d_state

        self.conv_states = []
        self.ssm_states = []
        self.transformer_layers = []
        for i in range(config.num_hidden_layers):
            if self.layers_block_type[i] == "mamba":
                self.conv_states += [
                    torch.zeros(
                        batch_size,
                        (config.mamba_expand * config.hidden_size + 2 * config.mamba_n_groups * ssm_state_size),
                        conv_kernel_size,
                        device=device,
                        dtype=dtype,
                    )
                ]
                self.ssm_states += [
                    torch.zeros(
                        batch_size,
                        config.mamba_n_heads,
                        config.mamba_d_head,
                        ssm_state_size,
                        device=device,
                        dtype=dtype,
                    )
                ]
            else:
                self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
                self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
                self.transformer_layers.append(i)

        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]

    def update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: Optional[dict[str, Any]] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Update the cache
        if self.key_cache[layer_idx].shape[-1] == 0:
            self.key_cache[layer_idx] = key_states
            self.value_cache[layer_idx] = value_states
        else:
            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)

        return self.key_cache[layer_idx], self.value_cache[layer_idx]

    def reorder_cache(self, beam_idx: torch.LongTensor):
        """Reorders the cache for beam search, given the selected beam indices."""
        for layer_idx in range(len(self.key_cache)):
            device = self.key_cache[layer_idx].device
            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
            device = self.value_cache[layer_idx].device
            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))

            device = self.conv_states[layer_idx].device
            self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
            device = self.ssm_states[layer_idx].device
            self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))

    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
        # take any layer that contains cache and not empty tensor
        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
        if len(self.key_cache) <= layer_idx:
            return 0
        return self.key_cache[layer_idx].shape[-2]

    def to_legacy_cache(self) -> tuple[tuple[torch.Tensor], tuple[torch.Tensor]]:
        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")

    @classmethod
    def from_legacy_cache(cls, past_key_values: Optional[tuple[tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")


# 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


is_fast_path_available = all((selective_state_update, causal_conv1d_fn, causal_conv1d_update))


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
    """
    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


# Adapted from transformers.models.mamba2.modeling_mamba2.Mamba2Mixer
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 HybridCache structure
    - There's a few non-obvious bugs fixed with batching in the slow path that exist in main
    - Some extra variables that our layer doesn't need have been removed
    - We ported most of the refactors in https://github.com/huggingface/transformers/pull/35154, which is (as of Dec 18, 2024) unmerged
    """

    def __init__(self, config: GraniteMoeHybridConfig, layer_idx: int):
        super().__init__()
        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.use_bias = config.mamba_proj_bias

        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

        # FIXME:
        self.time_step_limit = (0.0, float("inf"))
        self.time_step_min = 0.001
        self.time_step_max = 0.1

        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 of the input hidden states
        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)
        # instantiate once and copy inv_dt in init_weights of PretrainedModel
        self.dt_bias = nn.Parameter(torch.ones(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
        A = torch.arange(1, self.num_heads + 1)
        self.A_log = nn.Parameter(torch.log(A))
        self.A_log._no_weight_decay = True
        self.norm = GraniteMoeHybridRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon)
        self.D = nn.Parameter(torch.ones(self.num_heads))
        self.D._no_weight_decay = True

        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias)

        if not is_fast_path_available:
            logger.warning_once(
                "The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
                " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
                " https://github.com/Dao-AILab/causal-conv1d"
            )
        else:
            logger.warning_once("The fast path for GraniteMoeHybrid will be used when running the model on a GPU")

    def cuda_kernels_forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        seq_idx: Optional[torch.IntTensor] = None,
    ):
        # 1. Gated MLP's linear projection
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        projected_states = self.in_proj(hidden_states)

        # Set up dimensions for reshapes later
        batch_size, seq_len, _ = hidden_states.shape
        groups_time_state_size = self.n_groups * self.ssm_state_size

        use_precomputed_states = (
            cache_params is not None
            and cache_params.has_previous_state
            and seq_len == 1
            and cache_params.conv_states[self.layer_idx].shape[0]
            == cache_params.ssm_states[self.layer_idx].shape[0]
            == batch_size
            and cache_position is not None
            and cache_position[0] > 0
        )

        # getting projected states from cache if it exists
        if use_precomputed_states:
            gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
                [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
            )

            # 2. Convolution sequence transformation
            hidden_states_B_C = causal_conv1d_update(
                hidden_states_B_C,
                cache_params.conv_states[self.layer_idx],
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.activation,
            )

            hidden_states, B, C = torch.split(
                hidden_states_B_C,
                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                dim=-1,
            )

            # 3. SSM transformation
            A = -torch.exp(self.A_log.float())  # (nheads,)
            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
            D = self.D[:, None, ...].expand(-1, self.head_dim)
            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
            hidden_states = selective_state_update(
                cache_params.ssm_states[self.layer_idx],
                hidden_states_reshaped,
                dt,
                A,
                B,
                C,
                D,
                z=None,
                dt_bias=dt_bias,
                dt_softplus=True,
            )
            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
            hidden_states = self.norm(hidden_states, gate)

            # 4. Final linear projection
            out = self.out_proj(hidden_states)[:, None, ...]
        # Fused calculations or step by step if no initialized cache is found
        else:
            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}

            # 2-4. Fused kernel for conv1d, SSM, and the final projection
            if self.training and cache_params is None:
                out = mamba_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,
                    seq_idx=seq_idx,
                    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,
                    **dt_limit_kwargs,
                )

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

                # 2. Convolution sequence transformation
                # Init cache
                if cache_params is not None:
                    # storing the states
                    # If we just take xBC[:, :, -self.d_conv :], it will error if seqlen < self.d_conv
                    # Instead F.pad will pad with zeros if seqlen < self.d_conv, and truncate otherwise.
                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
                    conv_states = nn.functional.pad(
                        hidden_states_B_C_transposed,
                        (self.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
                    )
                    cache_params.conv_states[self.layer_idx].copy_(conv_states)

                if self.activation not in ["silu", "swish"]:
                    hidden_states_B_C = self.act(
                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
                    )
                else:
                    hidden_states_B_C = causal_conv1d_fn(
                        x=hidden_states_B_C.transpose(1, 2),
                        weight=self.conv1d.weight.squeeze(1),
                        bias=self.conv1d.bias,
                        activation=self.activation,
                        seq_idx=seq_idx,
                    ).transpose(1, 2)

                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
                hidden_states, B, C = torch.split(
                    hidden_states_B_C,
                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                    dim=-1,
                )

                # 3. SSM transformation
                scan_output, ssm_state = mamba_chunk_scan_combined(
                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
                    dt,
                    A,
                    B.view(batch_size, seq_len, self.n_groups, -1),
                    C.view(batch_size, seq_len, self.n_groups, -1),
                    chunk_size=self.chunk_size,
                    D=self.D,
                    z=None,
                    seq_idx=seq_idx,
                    return_final_states=True,
                    dt_bias=self.dt_bias,
                    dt_softplus=True,
                    **dt_limit_kwargs,
                )

                # Init cache
                if ssm_state is not None and cache_params is not None:
                    cache_params.ssm_states[self.layer_idx].copy_(ssm_state)

                scan_output = scan_output.view(batch_size, seq_len, -1)
                # Multiply "gate" branch and apply extra normalization layer
                scan_output = self.norm(scan_output, gate)

                # 4. Final linear projection
                out = self.out_proj(scan_output)
        return out

    # fmt: off
    def torch_forward(
        self,
        input_states,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ):
        batch_size, seq_len, _ = input_states.shape
        dtype = input_states.dtype

        # 1. Gated MLP's linear projection
        input_states = apply_mask_to_padding_states(input_states, attention_mask)
        projected_states = self.in_proj(input_states)
        gate, hidden_states_B_C, dt = projected_states.split(
                [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
        )

        use_precomputed_states = (
            cache_params is not None
            and cache_params.has_previous_state
            and seq_len == 1
            and cache_params.conv_states[self.layer_idx].shape[0]
            == cache_params.ssm_states[self.layer_idx].shape[0]
            == batch_size
            and cache_position is not None
            and cache_position[0] > 0
        )

        # 2. Convolution sequence transformation
        if use_precomputed_states:
            cache_params.conv_states[self.layer_idx] = cache_params.conv_states[self.layer_idx].roll(shifts=-1, dims=-1)
            cache_params.conv_states[self.layer_idx][:, :, -1] = hidden_states_B_C[:, 0, :].to(cache_params.conv_states[self.layer_idx].device)

            # We need to guarantee that anything regarding the cache is on the same device
            conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)

            hidden_states_B_C = torch.sum(
                conv_states * self.conv1d.weight.squeeze(1), dim=-1
            )
            if self.use_conv_bias:
                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
            hidden_states_B_C = self.act(hidden_states_B_C)
        else:
            # Init cache
            if cache_params is not None:
                hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
                conv_states = nn.functional.pad(
                    hidden_states_B_C_transposed, (self.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
                )
                cache_params.conv_states[self.layer_idx].copy_(conv_states)

            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))

        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, 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
        )

        # 3. SSM transformation
        A = -torch.exp(self.A_log.float())                            # [num_heads]
        if use_precomputed_states:
            # We need to guarantee that anything regarding the cache is on the same device
            cache_device = cache_params.ssm_states[self.layer_idx].device

            # Note: there is no need to pad parameter matrices here, as there is just one new token
            # for batched generation
            dt = dt[:, 0, :][:, None, ...]
            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
            # [num_heads] -> [num_heads, head_dim]
            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)

            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            # [bsz, num_heads, head_dim, state_size]
            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)

            # Discretize B
            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
            B = B.reshape(batch_size, -1, B.shape[-1])
            # [bsz, num_heads, head_dim, state_size]
            dB = dt[..., None] * B[..., None, :]

            # Discretize x into dB
            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
            dBx = (dB * hidden_states[..., None]).to(device=cache_device)

            # State calculation
            cache_params.ssm_states[self.layer_idx].copy_(
                cache_params.ssm_states[self.layer_idx] * dA + dBx
            )

            # Subsequent output
            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
            C = C.reshape(batch_size, -1, C.shape[-1])
            # [bsz, num_heads, head_dim]

            ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype)  # Shape: [b, h, d, n]
            # Reshape ssm_states to merge the first two dimensions
            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)  # Shape: [b*h, d, n]
            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
            y = torch.bmm(ssm_states_reshaped, C_reshaped)
            y = y.view(batch_size, self.num_heads, self.head_dim)

            # D skip connection
            # [num_heads] -> [num_heads, head_dim]
            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
            y = (y + hidden_states * D).to(y.dtype)

            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
            y = y.reshape(batch_size, -1)[:, None, ...]
        else:
            # begin ssd naive implementation without einsums
            dt = nn.functional.softplus(dt + self.dt_bias)
            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            B = B.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
            C = C.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size

            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)

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

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

            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
            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_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]  # shape: (b, c, l, s, h, n)
            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)

            # Compute M, equivalent to applying attention mask to weights
            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
            M = M_intermediate.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)
            if use_precomputed_states:
                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
            else:
                previous_states = torch.zeros_like(states[:, :1])
            states = torch.cat([previous_states, states], dim=1)
            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
            decay_chunk = decay_chunk.transpose(1, 3)
            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
            states, ssm_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, ...])
            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])

            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
            y = Y_diag + Y_off
            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)

            y = y + D_residual
            # Cutting off padded chunks
            if pad_size > 0:
                y = y[:, :seq_len, :, :]
            y = y.reshape(batch_size, seq_len, -1)

            # Init cache
            if ssm_state is not None and cache_params is not None:
                cache_params.ssm_states[self.layer_idx].copy_(ssm_state)

        scan_output = self.norm(y, gate)

        # end ssd naive

        # 4. Final linear projection
        contextualized_states = self.out_proj(scan_output.to(dtype))  # [batch, seq_len, hidden_size]
        return contextualized_states
    # fmt: on

    def forward(
        self,
        hidden_states,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        seq_idx: Optional[torch.IntTensor] = None,
        **kwargs,
    ):
        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask, seq_idx)
        if seq_idx is not None:
            raise NotImplementedError(
                "`seq_idx` support requires fast path support. Please install `mamba_ssm` and `causal_conv1d`"
            )
        dtype = hidden_states.dtype
        if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
            # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
            hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

        return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)


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 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.

    Attributes:
        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


class GraniteMoeHybridRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        GraniteMoeHybridRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        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 GraniteMoeHybridParallelExperts(nn.Module):
    def __init__(self, num_experts: int, input_size: int, output_size: int) -> None:
        """
        Initialize the GraniteMoeHybridParallelExperts module.
        The experts weights are stored in [num_experts, output_size, input_size] format. Such that it's compatible with
        many MoE libraries, such as [Megablock](https://github.com/databricks/megablocks) and
        [ScatterMoE](https://github.com/shawntan/scattermoe), as well as the
        [MoE kernel](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py)
        used in vllm.

        Args:
            num_experts (int):
                Number of experts.
            input_size (int):
                Size of the input.
            output_size (int):
                Size of the output.
        """
        super().__init__()
        self.weight = nn.Parameter(torch.empty(num_experts, output_size, input_size))
        self.num_experts = num_experts
        self.input_size = input_size
        self.output_size = output_size

    def forward(self, inputs, expert_size):
        """
        Forward pass of the GraniteMoeHybridParallelExperts module.

        Args:
            inputs (Tensor):
                Input tensor.
            expert_size:
                Expert size information.

        Returns:
            Tensor: Output tensor.
        """
        input_list = inputs.split(expert_size, dim=0)
        output_list = []
        for i in range(self.num_experts):
            output_list.append(F.linear(input_list[i], self.weight[i]))
        results = torch.cat(output_list, dim=0)
        return results


class GraniteMoeHybridTopKGating(nn.Module):
    def __init__(self, input_size: int, num_experts: int, top_k: int):
        """
        Initialize the top-k gating mechanism.
        Args:
            input_size (`int`):
                Size of the input.
            num_experts (`int`):
                Number of experts.
            top_k (`int`):
                Number of top experts to select.
        """
        super().__init__()

        self.num_experts = num_experts
        self.input_size = input_size
        self.top_k = top_k

        self.layer = nn.Linear(input_size, num_experts, bias=False)

    def forward(self, hidden_states):
        # compute the top_k routing decision
        logits = self.layer(hidden_states).float()  # [batch_size x seq_len, num_experts]
        top_k_logits, top_k_indices = logits.topk(self.top_k, dim=1)  # [num_tokens, top_k]
        top_k_gates = torch.softmax(top_k_logits, dim=1).type_as(hidden_states)  # [num_tokens, top_k]

        # compute number of input given to each expert
        zeros = torch.zeros(
            [top_k_gates.size(0), self.num_experts], dtype=top_k_gates.dtype, device=top_k_gates.device
        )  # [num_tokens, num_experts]
        gates = zeros.scatter(1, top_k_indices, 1)  # [num_tokens, num_experts]
        expert_size = gates.long().sum(0)  # [num_experts,]
        # (This cause torch.compile to fail with `torch._dynamo.exc.Unsupported: Backend compiler failed with a fake tensor exception at`)
        # (and `DataDependentOutputException`)
        expert_size = expert_size.tolist()

        # sort and group input tokens according to expert assignment
        top_k_experts = top_k_indices.flatten()  # [num_tokens * top_k]
        _, index_sorted_experts = top_k_experts.sort(0)  # [num_tokens * top_k]
        batch_index = index_sorted_experts.div(self.top_k, rounding_mode="trunc")  # [num_tokens * top_k]

        # gather the gate values for grouped input tokens
        top_k_gates = top_k_gates.flatten()  # [num_tokens * top_k]
        batch_gates = top_k_gates[index_sorted_experts]  # [num_tokens * top_k]

        return index_sorted_experts, batch_index, batch_gates, expert_size, logits


class GraniteMoeHybridMoE(nn.Module):
    """
    A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as 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.intermediate_size
        self.activation = ACT2FN[config.hidden_act]
        self.input_linear = GraniteMoeHybridParallelExperts(
            config.num_local_experts, self.input_size, self.hidden_size * 2
        )
        self.output_linear = GraniteMoeHybridParallelExperts(
            config.num_local_experts, self.hidden_size, self.input_size
        )

        self.router = GraniteMoeHybridTopKGating(
            input_size=self.input_size,
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
        )

    def forward(self, layer_input):
        """
        Forward pass of the mixture of experts layer.

        Args:
            layer_input (Tensor):
                Input tensor.

        Returns:
            Tensor:
                Output tensor.
            Tensor:
                Router logits.
        """
        bsz, length, emb_size = layer_input.size()
        layer_input = layer_input.reshape(-1, emb_size)
        _, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input)

        expert_inputs = layer_input[batch_index]
        hidden_states = self.input_linear(expert_inputs, expert_size)
        chunked_hidden_states = hidden_states.chunk(2, dim=-1)
        hidden_states = self.activation(chunked_hidden_states[0]) * chunked_hidden_states[1]
        expert_outputs = self.output_linear(hidden_states, expert_size)

        expert_outputs = expert_outputs * batch_gates[:, None]

        zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device)
        layer_output = zeros.index_add(0, batch_index, expert_outputs)
        layer_output = layer_output.view(bsz, length, self.input_size)
        return layer_output, router_logits


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
        if config.num_local_experts > 0:
            self.block_sparse_moe = GraniteMoeHybridMoE(config)
        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)

        self.residual_multiplier = config.residual_multiplier
        self.shared_mlp = GraniteMoeHybridMLP(config)
        self.mamba = None

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        output_router_logits: Optional[bool] = False,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs: Unpack[GraniteFlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*):
                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
                query_sequence_length, key_sequence_length)` if default attention is used.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs.Can be used to provide `GraniteFlashAttentionKwargs` for
                padding-free training and/or improve torch.compile performance.
        """
        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_position=cache_position,
                cache_params=past_key_value,
                attention_mask=attention_mask,
                **kwargs,
            )
            # No attention weights for state space layers
            self_attn_weights = None
        else:
            hidden_states, self_attn_weights = self.self_attn(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = residual + hidden_states * self.residual_multiplier

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)

        if self.has_experts:
            moe_hidden_states, router_logits = self.block_sparse_moe(hidden_states)
            hidden_states = moe_hidden_states + self.shared_mlp(hidden_states)
        else:
            hidden_states = self.shared_mlp(hidden_states)
            router_logits = None

        hidden_states = residual + hidden_states * self.residual_multiplier

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if output_router_logits:
            outputs += (router_logits,)

        return outputs


@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

    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
    _is_stateful = True

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, GraniteMoeHybridParallelExperts):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        if isinstance(module, GraniteMoeHybridMambaLayer):
            module.dt_bias.data.fill_(1.0)
            module.A_log.data = torch.log(torch.arange(1, module.num_heads + 1))
            module.D.data.fill_(1.0)
        elif isinstance(module, GraniteMoeHybridRMSNormGated):
            module.weight.data.fill_(1.0)


class GraniteMoeHybridRotaryEmbedding(nn.Module):
    def __init__(self, config: GraniteMoeHybridConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @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 torch.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)


@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.gradient_checkpointing = False

        self.embedding_multiplier = config.embedding_multiplier
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta

        self.position_embedding_type = config.position_embedding_type
        self.rotary_emb = GraniteMoeHybridRotaryEmbedding(config) if self.position_embedding_type == "rope" else None

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[GraniteFlashAttentionKwargs],
    ) -> Union[tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

        if self.gradient_checkpointing and self.training and use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
            )
            use_cache = False

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

        inputs_embeds = inputs_embeds * self.embedding_multiplier

        ## overwritten because `HybridMambaAttentionDynamicCache` is needed
        if use_cache and past_key_values is None:
            logger.warning_once(
                "GraniteMoeHybrid requires an initialized `HybridMambaAttentionDynamicCache` to return a cache. "
                "Because one was not provided, no cache will be returned."
            )

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )
        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )
        mamba_mask = self._update_mamba_mask(attention_mask, cache_position)

        # embed positions
        hidden_states = inputs_embeds

        position_embeddings = None
        # create position embeddings to be shared across the decoder layers
        if self.rotary_emb is not None:
            position_embeddings = self.rotary_emb(hidden_states, position_ids)

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

        for decoder_layer in self.layers:
            # Depending on the layer type we opt for 2D base attention mask (Mamba) or 4D causal mask (Attention)
            layer_mask = mamba_mask if decoder_layer.layer_type == "mamba" else causal_mask

            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=layer_mask,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                output_router_logits=output_router_logits,
                position_embeddings=position_embeddings,
                **kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                if layer_outputs[1] is not None:
                    # append attentions only of attention layers. Mamba layers return `None` as the attention weights
                    all_self_attns += (layer_outputs[1],)

            if output_router_logits:
                if layer_outputs[-1] is not None:
                    # append router logits only of expert layers. Regular MLP layers return `None` as the router logits
                    all_router_logits += (layer_outputs[-1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if past_key_values and not past_key_values.has_previous_state:
            past_key_values.has_previous_state = True

        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_logits=all_router_logits,
        )

    def _update_causal_mask(
        self,
        attention_mask: Union[torch.Tensor, "BlockMask"],
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool = False,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0.0).any():
                return attention_mask
            return None
        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)
            return attention_mask

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_compilable_cache = past_key_values.is_compileable if past_key_values is not None else False

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if self.config._attn_implementation == "sdpa" and not using_compilable_cache and not output_attentions:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype = input_tensor.dtype
        sequence_length = input_tensor.shape[1]
        if using_compilable_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type in ["cuda", "xpu", "npu"]
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            min_dtype = torch.finfo(dtype).min
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask

    def _update_mamba_mask(self, attention_mask, cache_position):
        """
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        """
        mamba_mask = attention_mask
        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
            mamba_mask = None
        return mamba_mask


def load_balancing_loss_func(
    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
    num_experts: Optional[int] = None,
    top_k=2,
    attention_mask: Optional[torch.Tensor] = None,
) -> Union[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


class GraniteMoeHybridForCausalLM(GraniteMoeHybridPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.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

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

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs,
    ) -> Union[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/PowerMoE-3b")
        >>> tokenizer = AutoTokenizer.from_pretrained("ibm/PowerMoE-3b")

        >>> 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_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # 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,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_router_logits=output_router_logits,
            return_dict=return_dict,
            cache_position=cache_position,
            **kwargs,
        )

        # Only compute necessary logits
        hidden_states = outputs[0]
        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:
            # Upcast to float if we need to compute the loss to avoid potential precision issues
            logits = logits.float()
            # 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 if return_dict else outputs[-1],
                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

        if not return_dict:
            output = (logits,) + outputs[1:]
            if output_router_logits:
                output = (aux_loss,) + output
            return (loss,) + output if loss is not None else output

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

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        **kwargs,
    ):
        # Overwritten -- has a unique cache type, `HybridMambaAttentionDynamicCache`

        empty_past_kv = past_key_values is None

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
        #              (we can't check exception 3 while compiling)
        if not empty_past_kv:
            if (
                inputs_embeds is not None  # Exception 1
                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
            ):
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]
        elif use_cache:
            past_key_values = HybridMambaAttentionDynamicCache(
                self.config, input_ids.shape[0], self.dtype, device=self.device
            )

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if not empty_past_kv:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and empty_past_kv:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
            }
        )
        return model_inputs


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