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

from collections.abc import Callable
from typing import Optional

import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
from ...masking_utils import create_bidirectional_mask, create_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
)
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, is_torchdynamo_compiling, logging
from ...utils.generic import maybe_autocast
from .configuration_dia import DiaConfig, DiaDecoderConfig, DiaEncoderConfig
from .generation_dia import DiaGenerationMixin


logger = logging.get_logger(__name__)


@auto_docstring
class DiaPreTrainedModel(PreTrainedModel):
    config: DiaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = True
    main_input_name = "input_ids"
    _no_split_modules = ["DiaEncoderLayer", "DiaDecoderLayer"]

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, DiaMultiChannelEmbedding):
            offsets = torch.arange(self.config.num_channels, dtype=torch.long) * self.config.vocab_size
            init.copy_(module.offsets, offsets)


class DiaMultiChannelEmbedding(nn.Module):
    """In order to efficiently compute the audio embedding from the 9 different channels,
    we vectorize the embedding process by using a single embedding layer and an offset.
    Example:
    - num_embeds = 4
    - vocab_size = 8
    - num_channels = 3
    We would have offsets = [0, 8, 16]
    If audio_codes = [0, 1, 2, 3], [1, 3, 4, 7], [5, 6, 7, 8],
    then tokens = audio_codes + offsets
                = [0, 1, 2, 3, 9, 11, 12, 15, 21, 22, 23, 24]
    This allows us to use a single embedding layer for all channels.
    """

    def __init__(self, config: DiaDecoderConfig):
        super().__init__()
        self.embed = nn.Embedding(config.vocab_size * config.num_channels, config.hidden_size)
        self.hidden_size = config.hidden_size
        self.num_channels = config.num_channels
        offsets = torch.arange(config.num_channels, dtype=torch.long) * config.vocab_size  # (C,)
        self.register_buffer("offsets", offsets, persistent=False)

    def forward(self, audio_codes: torch.Tensor) -> torch.Tensor:
        tokens = (audio_codes + self.offsets.to(audio_codes.device)).squeeze(1)
        embeds = self.embed(tokens).view(tokens.shape[0], audio_codes.shape[1], -1, self.hidden_size)
        return embeds.sum(dim=2)


class DiaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.config = config
        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
        self.activation_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
        up_states = self.gate_up_proj(hidden_states)

        gate, up_states = up_states.chunk(2, dim=-1)
        up_states = up_states * self.activation_fn(gate)

        return self.down_proj(up_states)


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

    def __init__(self, config: DiaConfig, 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: DiaConfig | 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)


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

    def __init__(self, config: DiaEncoderConfig | DiaDecoderConfig, layer_idx: int, is_causal: bool = False):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.hidden_size = config.hidden_size
        self.num_heads = self.config.num_attention_heads
        self.num_key_value_heads = self.config.num_key_value_heads or self.num_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.head_dim = getattr(config, "head_dim", config.hidden_size // self.num_heads)
        self.scaling = 1
        self.attention_dropout = 0.0
        self.is_causal = is_causal

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

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

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

        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:
            # 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_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 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


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

    def __init__(self, config: DiaDecoderConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.hidden_size = config.hidden_size
        self.cross_hidden_size = config.cross_hidden_size
        self.num_heads = self.config.cross_num_attention_heads
        self.num_key_value_heads = self.config.cross_num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.head_dim = config.cross_head_dim
        self.scaling = 1
        self.attention_dropout = 0.0
        self.is_causal = False

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        cross_attention_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        past_key_values: EncoderDecoderCache | 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)
        cross_shape = (*cross_attention_states.shape[:-1], -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        is_updated = past_key_values.is_updated.get(self.layer_idx) if past_key_values is not None else False
        if past_key_values is not None and is_updated:
            # reuse k,v, cross_attentions
            key_states = past_key_values.cross_attention_cache.layers[self.layer_idx].keys
            value_states = past_key_values.cross_attention_cache.layers[self.layer_idx].values
        else:
            key_states = self.k_proj(cross_attention_states).view(cross_shape).transpose(1, 2)
            value_states = self.v_proj(cross_attention_states).view(cross_shape).transpose(1, 2)

            if past_key_values is not None:
                # save all states to the cache
                key_states, value_states = past_key_values.cross_attention_cache.update(
                    key_states,
                    value_states,
                    self.layer_idx,
                )
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                past_key_values.is_updated[self.layer_idx] = True

        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,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape((*input_shape, -1)).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class DiaEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: DiaEncoderConfig, layer_idx: int):
        super().__init__()
        self.pre_sa_norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.self_attention = DiaSelfAttention(config, layer_idx, is_causal=False)
        self.post_sa_norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.mlp = DiaMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        residual = hidden_states
        normed_states = self.pre_sa_norm(hidden_states)
        self_attn_output, self_attn_weights = self.self_attention(
            normed_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = residual + self_attn_output

        residual = hidden_states
        normed_states = self.post_sa_norm(hidden_states)
        mlp_out = self.mlp(normed_states)
        hidden_states = residual + mlp_out

        return hidden_states, self_attn_weights


class DiaEncoder(DiaPreTrainedModel):
    def __init__(self, config: DiaEncoderConfig):
        super().__init__(config)
        self.config = config

        self.embedding = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList(
            [DiaEncoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.rotary_emb = DiaRotaryEmbedding(config=config)

        self.post_init()

    @auto_docstring
    @can_return_tuple
    def forward(
        self,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = False,
        output_hidden_states: bool | None = False,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> BaseModelOutput | tuple:
        hidden_states = self.embedding(input_ids)

        # RoPE
        # Note: We expect right padding and hence always generate
        # the position ids on the fly to reduce preparation overhead
        position_ids = torch.arange(input_ids.shape[-1], device=input_ids.device)[None, :]

        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=hidden_states,
            attention_mask=attention_mask,
        )
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)

            layer_outputs = encoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                position_embeddings=position_embeddings,
                **kwargs,
            )
            hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            encoder_states += (hidden_states,)

        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


class DiaDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: DiaDecoderConfig, layer_idx: int):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attention = DiaSelfAttention(config, layer_idx, is_causal=True)
        self.cross_attention = DiaCrossAttention(config, layer_idx)
        self.pre_sa_norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.pre_ca_norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.pre_mlp_norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.mlp = DiaMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor | None]:
        self_attn_cache = past_key_values
        if isinstance(self_attn_cache, EncoderDecoderCache):
            self_attn_cache = self_attn_cache.self_attention_cache

        residual = hidden_states
        normed_states = self.pre_sa_norm(hidden_states)
        self_attn_output, self_attn_weights = self.self_attention(
            normed_states,
            position_embeddings,
            attention_mask,
            # Needs to be an arg in order to function properly
            # on inplace operations to be carried (e.g. compile)
            self_attn_cache,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = residual + self_attn_output

        residual = hidden_states
        normed_states = self.pre_ca_norm(hidden_states)
        cross_states, cross_attn_weights = self.cross_attention(
            normed_states,
            encoder_hidden_states,
            attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            **kwargs,
        )
        hidden_states = residual + cross_states

        residual = hidden_states
        normed_states = self.pre_mlp_norm(hidden_states)
        mlp_out = self.mlp(normed_states)
        hidden_states = residual + mlp_out

        return hidden_states, self_attn_weights, cross_attn_weights


class DiaDecoder(DiaPreTrainedModel):
    """Transformer Decoder Stack using DenseGeneral."""

    def __init__(self, config: DiaDecoderConfig):
        super().__init__(config)
        self.num_channels = config.num_channels
        self.vocab_size = config.vocab_size
        self.embeddings = DiaMultiChannelEmbedding(config)
        self.layers = nn.ModuleList(
            [DiaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = DiaRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.rotary_emb = DiaRotaryEmbedding(config=config)

        self.post_init()

    @auto_docstring
    @can_return_tuple
    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.LongTensor | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        output_attentions: bool | None = False,
        output_hidden_states: bool | None = False,
        cache_position: torch.LongTensor | None = None,
        **kwargs,
    ) -> BaseModelOutputWithPastAndCrossAttentions | tuple:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, num_codebooks)`):
            The original `decoder_input_ids` in 3D shape to facilitate more efficient computations.

            [What are input IDs?](../glossary#input-ids)
        """

        batch_size, seq_length = input_ids.size()[:-1]
        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        if cache_position is None:
            cache_position = torch.arange(
                past_key_values_length, past_key_values_length + seq_length, device=input_ids.device
            )
        if position_ids is None:
            position_ids = cache_position[None, :]

        # RoPE
        hidden_states = self.embeddings(input_ids)

        if attention_mask is None and not is_torchdynamo_compiling():
            # required mask seq length can be calculated via length of past cache
            mask_seq_length = past_key_values_length + seq_length
            attention_mask = torch.ones(batch_size, mask_seq_length, device=input_ids.device)

        attention_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=hidden_states,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
        )
        encoder_attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=hidden_states,
            attention_mask=encoder_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
        )
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None

        for layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = layer(
                hidden_states,
                # Needs to be an arg in order to function properly
                # on inplace operations to be carried (e.g. compile)
                position_embeddings,
                attention_mask,
                encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                past_key_values=past_key_values,
                cache_position=cache_position,
                position_ids=position_ids,
                **kwargs,
            )
            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns = all_self_attns + (layer_outputs[1],)

                if encoder_hidden_states is not None:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )


@auto_docstring(
    custom_intro="""
    The bare Dia model outputting raw hidden-states without any specific head on top.
    """
)
class DiaModel(DiaPreTrainedModel):
    def __init__(self, config: DiaConfig):
        super().__init__(config)
        self.config = config
        self.encoder = DiaEncoder(config.encoder_config)
        self.decoder = DiaDecoder(config.decoder_config)
        self.post_init()

    @auto_docstring
    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.LongTensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_position_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.LongTensor | None = None,
        encoder_outputs: BaseModelOutput | tuple | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs,
    ) -> tuple | Seq2SeqModelOutput:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size * num_codebooks, target_sequence_length)
        or (batch_size, target_sequence_length, num_codebooks)`, *optional*):
            1. (batch_size * num_codebooks, target_sequence_length): corresponds to the general use case where
            the audio input codebooks are flattened into the batch dimension. This also aligns with the flat-
            tened audio logits which are used to calculate the loss.

            2. (batch_size, sequence_length, num_codebooks): corresponds to the internally used shape of
            Dia to calculate embeddings and subsequent steps more efficiently.

            If no `decoder_input_ids` are provided, it will create a tensor of `bos_token_id` with shape
            `(batch_size, 1, num_codebooks)`. Indices can be obtained using the [`DiaProcessor`]. See
            [`DiaProcessor.__call__`] for more details.

            [What are decoder input IDs?](../glossary#decoder-input-ids)
        decoder_position_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`):
            Indices of positions of each input sequence tokens in the position embeddings.
            Used to calculate the position embeddings up to `config.decoder_config.max_position_embeddings`.

            [What are position IDs?](../glossary#position-ids)
        """

        if input_ids is None and encoder_outputs is None:
            raise ValueError(
                "You should either provide text ids or the cached text encodings. Neither has been found."
            )

        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

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

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

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                **kwargs,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput
        elif not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        # On default we initialize the decoder with bos tokens if nothing has been provided
        bsz, seq_len, channels = (encoder_outputs[0].shape[0], -1, self.config.decoder_config.num_channels)
        if decoder_input_ids is None:
            decoder_input_ids = torch.full(
                size=(bsz, 1, channels), fill_value=self.config.decoder_config.bos_token_id, device=self.device
            )
        # Ensure 3D
        if decoder_input_ids.ndim == 2:
            decoder_input_ids = decoder_input_ids.reshape(bsz, channels, seq_len).transpose(1, 2)

        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            position_ids=decoder_position_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            use_cache=use_cache,
            cache_position=cache_position,
            **kwargs,
        )

        return Seq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs[0],
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    The Dia model consisting of a (byte) text encoder and audio decoder with a prediction head on top.
    """
)
class DiaForConditionalGeneration(DiaPreTrainedModel, DiaGenerationMixin):
    base_model_prefix = "model"
    output_modalities = ("audio",)

    def __init__(self, config: DiaConfig):
        super().__init__(config)
        self.config = config
        self.model = DiaModel(config)

        self.num_channels = config.decoder_config.num_channels
        self.vocab_size = config.decoder_config.vocab_size
        self.logits_dense = nn.Linear(
            config.decoder_config.hidden_size, (self.num_channels * self.vocab_size), bias=False
        )
        self.loss_type = "ForMaskedLM"

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

    @auto_docstring
    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.LongTensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_position_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.LongTensor | None = None,
        encoder_outputs: BaseModelOutput | tuple | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        labels: torch.LongTensor | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs,
    ) -> tuple | Seq2SeqLMOutput:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size * num_codebooks, target_sequence_length)
        or (batch_size, target_sequence_length, num_codebooks)`, *optional*):
            1. (batch_size * num_codebooks, target_sequence_length): corresponds to the general use case where
            the audio input codebooks are flattened into the batch dimension. This also aligns with the flat-
            tened audio logits which are used to calculate the loss.

            2. (batch_size, sequence_length, num_codebooks): corresponds to the internally used shape of
            Dia to calculate embeddings and subsequent steps more efficiently.

            If no `decoder_input_ids` are provided, it will create a tensor of `bos_token_id` with shape
            `(batch_size, 1, num_codebooks)`. Indices can be obtained using the [`DiaProcessor`]. See
            [`DiaProcessor.__call__`] for more details.

            [What are decoder input IDs?](../glossary#decoder-input-ids)
        decoder_position_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`):
            Indices of positions of each input sequence tokens in the position embeddings.
            Used to calculate the position embeddings up to `config.decoder_config.max_position_embeddings`.

            [What are position IDs?](../glossary#position-ids)
        labels (`torch.LongTensor` of shape `(batch_size * num_codebooks,)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in
            `[0, ..., config.decoder_config.vocab_size - 1]` or -100. Tokens with indices set to `-100`
            are ignored (masked).
        """

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_position_ids=decoder_position_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            cache_position=cache_position,
            **kwargs,
        )

        last_hidden_state = outputs[0]
        batch_size = last_hidden_state.shape[0]
        # 3D <-> 2D makes it necessary to prioritize channel dim
        audio_logits = (
            self.logits_dense(last_hidden_state)
            .view((batch_size, -1, self.num_channels, self.vocab_size))
            .transpose(1, 2)
            .contiguous()
            .view(batch_size * self.num_channels, -1, self.vocab_size)
        )

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

        return Seq2SeqLMOutput(
            loss=loss,
            logits=audio_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )


__all__ = ["DiaModel", "DiaPreTrainedModel", "DiaForConditionalGeneration"]