#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/beam/modular_beam.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_beam.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 Reflection AI 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.

import math
from collections.abc import Callable

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernelized_func
from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, slice_logits_to_keep
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_beam import BeamConfig


@use_kernel_forward_from_hub("RMSNorm")
class BeamRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        BeamRMSNorm 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:
        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 BeamRotaryEmbedding(nn.Module):
    def __init__(self, config: BeamConfig, 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.layer_types = sorted(set(config.layer_types))
        self.rope_type = {}
        for layer_type in self.layer_types:
            rope_params = self.config.rope_parameters[layer_type]
            if rope_params is None:
                continue

            self.rope_type[layer_type] = rope_params["rope_type"]
            rope_init_fn: Callable = self.compute_default_rope_parameters
            if self.rope_type[layer_type] != "default":
                rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
            curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, device, layer_type=layer_type)
            setattr(self, f"{layer_type}_inv_freq", nn.Buffer(curr_inv_freq, persistent=False))
            setattr(self, f"{layer_type}_original_inv_freq", nn.Buffer(curr_inv_freq.clone(), persistent=False))
            setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)

    @staticmethod
    def compute_default_rope_parameters(
        config: BeamConfig, device=None, layer_type: str | None = None, **kwargs
    ) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            layer_type (`str`):
                The current layer type if the model has different RoPE parameters per type.
                Should not be used unless `config.layer_types is not None`

        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        # For backward compatibility standardize the `rope_parameters_dict` if it uses old format
        base = config.rope_parameters[layer_type]["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

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

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids, layer_type):
        # diff -> returns cos/sin in fp32 without casting to `x.dtype`
        inv_freq = getattr(self, f"{layer_type}_inv_freq")
        attention_scaling = getattr(self, f"{layer_type}_attention_scaling")

        freqs = position_ids[..., None].float() * inv_freq.to(device=x.device, dtype=torch.float)
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos() * attention_scaling
        sin = emb.sin() * attention_scaling

        return cos, sin


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


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

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


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

    def __init__(self, config: BeamConfig, 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.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.q_norm = BeamRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
        self.k_norm = BeamRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None

        self.gate_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)
        # One gate bias per head, broadcast over `head_dim`.
        self.gate_bias = nn.Parameter(torch.zeros(config.num_attention_heads))

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

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

        # Sliding-attention layers use RoPE; full-attention layers use NoPE with position-dependent query scaling.
        if position_embeddings is not None:
            cos, sin = position_embeddings
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
        if attention_scaling is not None:
            query_states = (query_states.float() * attention_scaling).to(query_states.dtype)

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

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

        gate = self.gate_proj(hidden_states).view_as(attn_output) + self.gate_bias[None, None, :, None]
        attn_output = attn_output * torch.sigmoid(gate)
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


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

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

    def forward(
        self, hidden_states: torch.Tensor, top_k_index: torch.Tensor, top_k_weights: torch.Tensor
    ) -> torch.Tensor:
        final = torch.zeros_like(hidden_states)
        with torch.no_grad():
            mask = F.one_hot(top_k_index, num_classes=self.num_experts + 1).permute(2, 1, 0)
            hit = torch.greater(mask.sum(dim=(-1, -2)), 0).nonzero()
        for expert_idx in hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(mask[expert_idx])
            current = self._apply_gate(F.linear(hidden_states[token_idx], self.gate_up_proj[expert_idx]))
            current = F.linear(current, self.down_proj[expert_idx]) * top_k_weights[token_idx, top_k_pos, None]
            final.index_add_(0, token_idx, current.to(final.dtype))
        return final

    def _apply_gate(self, gate_up: torch.Tensor) -> torch.Tensor:
        gate, up = gate_up.chunk(2, dim=-1)
        return (self.act_fn(gate.float()) * up.float()).to(gate.dtype)


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

    def forward(self, x):
        gate = self.gate_proj(x)
        up = self.up_proj(x)
        # Beam evaluates the SwiGLU activation and product in FP32.
        activated = (self.act_fn(gate.float()) * up.float()).to(gate.dtype)
        return self.down_proj(activated)


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

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


class BeamMoE(nn.Module):
    def __init__(self, config: BeamConfig):
        super().__init__()
        self.experts = BeamExperts(config)
        self.gate = BeamTopKRouter(config)
        self.routed_scaling_factor = config.routed_scaling_factor

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states_reshaped = hidden_states.view(-1, hidden_dim)
        _, routing_weights, selected_experts = self.gate(hidden_states_reshaped)
        final_hidden_states = self.experts(hidden_states_reshaped, selected_experts, routing_weights)
        final_hidden_states = final_hidden_states * self.routed_scaling_factor
        return final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)


def depth_scaled_residual_add(
    hidden_states: torch.Tensor,
    residual: torch.Tensor,
    residual_add_idx: int,
) -> torch.Tensor:
    hidden_states_scale = 1.0 / math.sqrt(residual_add_idx)
    residual_scale = math.sqrt((residual_add_idx - 1) / residual_add_idx)
    return hidden_states * hidden_states_scale + residual * residual_scale


class BeamDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: BeamConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = BeamAttention(config, layer_idx)

        self.mlp = BeamMoE(config) if config.mlp_layer_types[layer_idx] == "sparse" else BeamMLP(config)

        self.input_layernorm = BeamRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.post_attention_layernorm = BeamRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.pre_feedforward_layernorm = BeamRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.post_feedforward_layernorm = BeamRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.attn_residual_add_idx = 2 * (layer_idx + 1)
        self.mlp_residual_add_idx = self.attn_residual_add_idx + 1

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        attention_scaling: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            attention_scaling=attention_scaling,
            past_key_values=past_key_values,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = depth_scaled_residual_add(hidden_states, residual, self.attn_residual_add_idx)

        residual = hidden_states
        hidden_states = self.pre_feedforward_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = depth_scaled_residual_add(hidden_states, residual, self.mlp_residual_add_idx)
        return hidden_states


@auto_docstring
class BeamPreTrainedModel(PreTrainedModel):
    config: BeamConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["BeamDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "router_logits": OutputRecorder(BeamTopKRouter, index=0),
        "hidden_states": BeamDecoderLayer,
        "attentions": BeamAttention,
    }
    _keep_in_fp32_modules_strict = ["e_score_correction_bias", "mlp.gate.weight"]

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, BeamTopKRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.e_score_correction_bias)
        elif isinstance(module, BeamExperts):
            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, BeamRotaryEmbedding):
            for layer_type in module.rope_type:
                rope_init_fn = module.compute_default_rope_parameters
                if module.rope_type[layer_type] != "default":
                    rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
                curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
                init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
                init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)
        elif isinstance(module, BeamAttention):
            init.zeros_(module.gate_bias)


def compute_attention_scaling(position_ids: torch.LongTensor) -> torch.Tensor:
    """Scalable-softmax query scale, `(1 + position / 2048) ** 0.1`, for layers that run without RoPE."""
    return torch.exp(0.1 * torch.log1p((position_ids.float() + 1) / 2048))[:, None, :, None]


@auto_docstring
class BeamModel(BeamPreTrainedModel):
    def __init__(self, config: BeamConfig):
        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(
            [BeamDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = BeamRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = BeamRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

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

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

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

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

        if not isinstance(causal_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            mask_creation_functions = {
                "full_attention": lambda: create_causal_mask(**mask_kwargs),
                "sliding_attention": lambda: create_sliding_window_causal_mask(**mask_kwargs),
            }
            causal_mask_mapping = {}
            for layer_type in set(self.config.layer_types):
                causal_mask_mapping[layer_type] = mask_creation_functions[layer_type]()

        hidden_states = inputs_embeds
        position_embeddings, attention_scalings = {}, {}
        for layer_type in set(self.config.layer_types):
            if self.config.rope_parameters[layer_type] is None:
                # Full-attention layers use NoPE with position-dependent attention scaling.
                position_embeddings[layer_type] = None
                attention_scalings[layer_type] = compute_attention_scaling(position_ids)
            else:
                position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type)
                attention_scalings[layer_type] = None

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

        hidden_states = self.norm(hidden_states)
        return MoeModelOutputWithPast(last_hidden_state=hidden_states, past_key_values=past_key_values)


@auto_docstring
class BeamForCausalLM(BeamPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    _fsdp_plan = {"lm_head": "keep_full_weight"}

    def __init__(self, config: BeamConfig):
        super().__init__(config)
        self.model = BeamModel(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.Tensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> 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]`.
        """
        outputs: MoeModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        hidden_states = slice_logits_to_keep(hidden_states, logits_to_keep)
        logits = self.lm_head(hidden_states)
        if self.config.final_logit_softcapping is not None:
            # Softcap in float32, as bf16 tanh quantizes badly near saturation.
            logits = logits.float() / self.config.final_logit_softcapping
            logits = torch.tanh(logits)
            logits = logits * self.config.final_logit_softcapping

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

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


__all__ = ["BeamForCausalLM", "BeamModel", "BeamPreTrainedModel"]
