#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/lfm2_moe/modular_lfm2_moe.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_lfm2_moe.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

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

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

from ... import initialization as init
from ...cache_utils import Cache
from ...generation import GenerationMixin
from ...integrations import (
    use_experts_implementation,
    use_kernel_forward_from_hub,
    use_kernel_func_from_hub,
    use_kernelized_func,
)
from ...masking_utils import create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, MoeModelOutputWithPast
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.import_utils import is_causal_conv1d_available, is_torchdynamo_compiling
from ...utils.output_capturing import capture_outputs
from .configuration_lfm2_moe import Lfm2MoeConfig


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


@use_kernel_forward_from_hub("RMSNorm")
class Lfm2MoeRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        Lfm2MoeRMSNorm 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 Lfm2MoeRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Lfm2MoeConfig, device=None):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_default_rope_parameters
        if self.rope_type != "default":
            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)

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

    @staticmethod
    def compute_default_rope_parameters(
        config: Lfm2MoeConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

        attention_factor = 1.0  # Unused in this type of RoPE

        # Compute the inverse frequencies
        inv_freq = 1.0 / (
            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
        )
        return inv_freq, attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        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 Lfm2MoeMLP(nn.Module):
    def __init__(self, config: Lfm2MoeConfig, intermediate_size: int | None = None):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
        self.w1 = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.w3 = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.w2 = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)

    def forward(self, x):
        return self.w2(F.silu(self.w1(x)) * self.w3(x))


@use_experts_implementation
class Lfm2MoeExperts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

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

    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 Lfm2MoeSparseMoeBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.routed_scaling_factor = config.routed_scaling_factor
        self.norm_topk_prob = config.norm_topk_prob
        self.use_expert_bias = config.use_expert_bias

        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.experts = Lfm2MoeExperts(config)
        if self.use_expert_bias:
            self.register_buffer("expert_bias", torch.zeros(config.num_experts, dtype=torch.float32))

    def route_tokens_to_experts(self, router_logits):
        routing_weights = router_logits.sigmoid()
        if self.use_expert_bias:
            scores_for_routing = routing_weights + self.expert_bias
            _, selected_experts = torch.topk(scores_for_routing, k=self.top_k, dim=-1)
            routing_weights = torch.gather(routing_weights, dim=1, index=selected_experts).type_as(router_logits)
        else:
            routing_weights, selected_experts = torch.topk(routing_weights, k=self.top_k, dim=-1)

        if self.norm_topk_prob:
            routing_weights = routing_weights / (routing_weights.sum(dim=-1, keepdim=True) + 1e-6)
        routing_weights = routing_weights * self.routed_scaling_factor
        return selected_experts, routing_weights

    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states_reshaped = hidden_states.view(-1, hidden_dim)
        router_logits = self.gate(hidden_states_reshaped)
        selected_experts, routing_weights = self.route_tokens_to_experts(router_logits)
        final_hidden_states = self.experts(hidden_states_reshaped, selected_experts, routing_weights)
        return final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)


class Lfm2MoeHybridConvCache:
    """
    Attention and conv cache for Lfm2Moe.

    It stores the Key and Value states as a list of tensors, one for each layer.
    Attention layer cache shape: `[batch_size, num_heads, seq_len, head_dim]`.
    Conv layer cache shape: `[batch_size, hidden_size, L_cache-1]`.
    """

    # Override @property existing in Cache
    max_batch_size = None
    is_compileable = False
    key_cache = None
    value_cache = None

    def __init__(
        self,
        config: Lfm2MoeConfig,
        max_batch_size: int,
        dtype: torch.dtype = torch.float32,
        device: torch.device | str | None = None,
    ):
        self.key_cache = []
        self.value_cache = []
        self.max_batch_size = max_batch_size
        self.layer_types = config.layer_types
        self.first_attention_layer = self.layer_types.index("full_attention")
        self.conv_L_cache = config.conv_L_cache
        self._dtype = dtype

        self.conv_cache: list[torch.Tensor] = []
        device = torch.device(device) if device is not None else None

        for _ in range(config.num_hidden_layers):
            conv_state = torch.zeros(
                self.max_batch_size,
                config.hidden_size,
                self.conv_L_cache,
                dtype=self._dtype,
                device=device,
            )
            self.conv_cache.append(conv_state)
            self.key_cache.append(torch.tensor([]))
            self.value_cache.append(torch.tensor([]))

    def update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: dict[str, Any] | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.

        Parameters:
            key_states (`torch.Tensor`):
                The new key states to cache.
            value_states (`torch.Tensor`):
                The new value states to cache.
            layer_idx (`int`):
                The index of the layer to cache the states for.
            cache_kwargs (`Dict[str, Any]`, `optional`):
                Additional arguments for the cache subclass. No additional arguments are used in `DynamicCache`.

        Return:
            A tuple containing the updated key and value states.
        """
        # Update the cache
        if self.key_cache[layer_idx].numel() == 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)):
            if self.key_cache[layer_idx].numel():
                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))

            if self.conv_cache[layer_idx].numel():
                device = self.conv_cache[layer_idx].device
                self.conv_cache[layer_idx] = self.conv_cache[layer_idx].index_select(0, beam_idx.to(device))

    def get_seq_length(self, layer_idx: int | None = 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.first_attention_layer if self.layer_types[layer_idx] != "full_attention" else layer_idx
        if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx].numel() == 0:
            return 0
        return self.key_cache[layer_idx].shape[-2]

    def get_mask_sizes(self, cache_position: torch.Tensor, layer_idx: int) -> tuple[int, int]:
        """
        Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for
        the given layer at `layer_idx`.
        The masks are then prepared according to the given lengths (kv_length, kv_offset) and patterns (i.e. sliding_window, chunk_size),
        for each layer.
        """
        full_mask_kv_offset = 0
        query_length = cache_position.shape[0]
        past_seen_tokens = self.get_seq_length()
        kv_length = query_length + past_seen_tokens
        return kv_length, full_mask_kv_offset

    def crop(self, max_length: int):
        """Crop the cache to the given length"""
        if max_length < 0:
            max_length = self.get_seq_length() - abs(max_length)

        if self.get_seq_length() <= max_length:
            return

        for idx in range(len(self.key_cache)):
            if self.key_cache[idx].numel():
                self.key_cache[idx] = self.key_cache[idx][..., :max_length, :]
                self.value_cache[idx] = self.value_cache[idx][..., :max_length, :]

    def __len__(self) -> int:
        return len(self.key_cache)

    def reset(self):
        for layer_idx in range(len(self.conv_cache)):
            # In-place ops prevent breaking the static address
            self.conv_cache[layer_idx].zero_()


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_func_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 Lfm2MoeAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Lfm2MoeConfig, 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 = self.head_dim**-0.5
        self.is_causal = True
        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.out_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
        self.q_layernorm = Lfm2MoeRMSNorm(self.head_dim, eps=config.norm_eps)
        self.k_layernorm = Lfm2MoeRMSNorm(self.head_dim, eps=config.norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        past_key_values: Lfm2MoeHybridConvCache | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

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

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

        if past_key_values is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        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,
            scaling=self.scaling,
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        output = self.out_proj(attn_output)
        return output, attn_weights


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 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


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


class Lfm2MoeShortConv(nn.Module):
    def __init__(
        self,
        config: Lfm2MoeConfig,
        layer_idx: int,
    ):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.L_cache = config.conv_L_cache
        self.bias = config.conv_bias

        self.conv = nn.Conv1d(
            in_channels=config.hidden_size,
            out_channels=config.hidden_size,
            kernel_size=self.L_cache,
            groups=config.hidden_size,
            bias=self.bias,
            padding=self.L_cache - 1,
        )
        self.in_proj = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=self.bias)
        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=self.bias)

    def cuda_kernels_forward(
        self,
        x: torch.Tensor,
        past_key_values: Lfm2MoeHybridConvCache | None = None,
        cache_position: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
    ):
        x = apply_mask_to_padding_states(x, attention_mask)
        BCx = self.in_proj(x).transpose(-1, -2)
        B, C, x = BCx.chunk(3, dim=-2)

        Bx = B * x

        conv_weights = self.conv.weight.view(self.conv.weight.size(0), self.conv.weight.size(2))
        if past_key_values is not None and cache_position[0] > 0:
            conv_out = causal_conv1d_update(
                Bx.squeeze(-1),
                past_key_values.conv_cache[self.layer_idx],
                conv_weights,
                self.conv.bias,
                None,
            )
            conv_out = conv_out.unsqueeze(-1)
        else:
            if past_key_values is not None:
                conv_state = nn.functional.pad(Bx, (self.L_cache - Bx.shape[-1], 0))
                past_key_values.conv_cache[self.layer_idx].copy_(conv_state)

            conv_out = causal_conv1d_fn(Bx, conv_weights, self.conv.bias, activation=None)

        y = C * conv_out
        y = self.out_proj(y.transpose(-1, -2).contiguous())
        return y

    def slow_forward(
        self,
        x: torch.Tensor,
        past_key_values: Lfm2MoeHybridConvCache | None = None,
        cache_position: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
    ):
        seqlen = x.shape[1]

        x = apply_mask_to_padding_states(x, attention_mask)
        BCx = self.in_proj(x).transpose(-1, -2)
        B, C, x = BCx.chunk(3, dim=-2)

        Bx = B * x

        if past_key_values is not None and cache_position[0] > 0:
            conv_state = past_key_values.conv_cache[self.layer_idx]
            cache_position = cache_position.clamp(0, self.L_cache - 1)
            conv_state = conv_state.roll(shifts=-1, dims=-1)
            conv_state[:, :, cache_position] = Bx.to(device=conv_state.device, dtype=conv_state.dtype)
            past_key_values.conv_cache[self.layer_idx].copy_(conv_state)
            conv_out = torch.sum(conv_state.to(Bx.device) * self.conv.weight[:, 0, :], dim=-1)
            if self.bias:
                conv_out += self.conv.bias

            conv_out = conv_out.unsqueeze(-1)
        else:
            if past_key_values is not None:
                conv_state = nn.functional.pad(Bx, (self.L_cache - Bx.shape[-1], 0))
                past_key_values.conv_cache[self.layer_idx].copy_(conv_state)

            conv_out = self.conv(Bx)[..., :seqlen]

        y = C * conv_out
        y = y.transpose(-1, -2).contiguous()
        y = self.out_proj(y)
        return y

    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Lfm2MoeHybridConvCache | None = None,
        cache_position: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
    ):
        if is_fast_path_available and "cuda" in hidden_states.device.type and not is_torchdynamo_compiling():
            return self.cuda_kernels_forward(hidden_states, past_key_values, cache_position, attention_mask)
        return self.slow_forward(hidden_states, past_key_values, cache_position, attention_mask)


class Lfm2MoeDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Lfm2MoeConfig, layer_idx: int):
        super().__init__()
        self.is_attention_layer = config.layer_types[layer_idx] == "full_attention"

        if self.is_attention_layer:
            self.self_attn = Lfm2MoeAttention(config, layer_idx)
        else:
            self.conv = Lfm2MoeShortConv(config, layer_idx)
        self.feed_forward = (
            Lfm2MoeMLP(config, intermediate_size=config.intermediate_size)
            if layer_idx < config.num_dense_layers
            else Lfm2MoeSparseMoeBlock(config)
        )
        self.operator_norm = Lfm2MoeRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.ffn_norm = Lfm2MoeRMSNorm(config.hidden_size, eps=config.norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Lfm2MoeHybridConvCache | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs,
    ) -> torch.Tensor:
        residual = hidden_states
        if self.is_attention_layer:
            hidden_states, _ = self.self_attn(
                hidden_states=self.operator_norm(hidden_states),
                position_embeddings=position_embeddings,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                cache_position=cache_position,
                **kwargs,
            )
        else:
            hidden_states = self.conv(
                hidden_states=self.operator_norm(hidden_states),
                past_key_values=past_key_values,
                cache_position=cache_position,
                attention_mask=attention_mask,
            )
        hidden_states = hidden_states + residual
        hidden_states = hidden_states + self.feed_forward(self.ffn_norm(hidden_states))

        return hidden_states


@auto_docstring
class Lfm2MoePreTrainedModel(PreTrainedModel):
    config: Lfm2MoeConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Lfm2MoeDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False  # uses a non-compilable custom cache class Lfm2MoeHybridConvCache
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Lfm2MoeDecoderLayer,
        "attentions": Lfm2MoeAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Lfm2MoeExperts):
            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, Lfm2MoeSparseMoeBlock):
            if module.use_expert_bias:
                init.zeros_(module.expert_bias)


@auto_docstring
class Lfm2MoeModel(Lfm2MoePreTrainedModel):
    def __init__(self, config: Lfm2MoeConfig):
        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(
            [Lfm2MoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.gradient_checkpointing = False
        self.pos_emb = Lfm2MoeRotaryEmbedding(config)
        self.embedding_norm = Lfm2MoeRMSNorm(config.hidden_size, eps=config.norm_eps)

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Lfm2MoeHybridConvCache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

        if use_cache and past_key_values is None:
            batch_size = inputs_embeds.shape[0]
            past_key_values = Lfm2MoeHybridConvCache(
                config=self.config, max_batch_size=batch_size, dtype=self.dtype, device=self.device
            )

        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 = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )
        # Skip masking for decoding stage. We check shape here to be compile-friendly
        linear_attention = attention_mask if inputs_embeds.shape[1] != 1 else None

        hidden_states = inputs_embeds
        position_embeddings = self.pos_emb(hidden_states, position_ids=position_ids)

        # decoder layers
        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            layer_mask = causal_mask if decoder_layer.is_attention_layer else linear_attention
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=layer_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.embedding_norm(hidden_states)

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


@auto_docstring
class Lfm2MoeForCausalLM(Lfm2MoePreTrainedModel, 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"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = Lfm2MoeModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from transformers import AutoTokenizer, Lfm2MoeForCausalLM

        >>> model = Lfm2MoeForCausalLM.from_pretrained("meta-lfm2_moe/Lfm2Moe-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-lfm2_moe/Lfm2Moe-2-7b-hf")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        outputs: BaseModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            cache_position=cache_position,
            **kwargs,
        )

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

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

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


__all__ = ["Lfm2MoeForCausalLM", "Lfm2MoeModel", "Lfm2MoePreTrainedModel"]