layers.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from huggingface_hub import PyTorchModelHubMixin
from torch import Tensor
from torch.nn import RMSNorm

from .config import DecoderConfig, DiaConfig, EncoderConfig
from .state import DecoderInferenceState, EncoderInferenceState, KVCache


def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
    return tuple(ax if ax >= 0 else ndim + ax for ax in axes)


class DenseGeneral(nn.Module):
    """
    PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
    Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
    for the generalized matrix multiplication. Weight/bias shapes are calculated
    and parameters created during initialization based on config.
    `load_weights` validates shapes and copies data.
    Attributes:
        axis (Tuple[int, ...]): Input axis or axes to contract.
        in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
        out_features (Tuple[int, ...]): Shape of the output features (non-contracted dims).
        use_bias (bool): Whether to add a bias term.
        weight (nn.Parameter): The kernel parameter.
        bias (Optional[nn.Parameter]): The bias parameter (if use_bias=True).
    """

    def __init__(
        self,
        in_shapes: tuple[int, ...],
        out_features: tuple[int, ...],
        axis: tuple[int, ...] = (-1,),
        weight_dtype: torch.dtype | None = None,
        device: torch.device | None = None,
    ):
        super().__init__()
        self.in_shapes = in_shapes
        self.out_features = out_features
        self.axis = axis
        self.kernel_shape = self.in_shapes + self.out_features

        factory_kwargs = {"device": device, "dtype": weight_dtype}
        self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))

    def forward(self, inputs: Tensor) -> Tensor:
        norm_axis = _normalize_axes(self.axis, inputs.ndim)
        kernel_contract_axes = tuple(range(len(norm_axis)))

        output = torch.tensordot(
            inputs.to(self.weight.dtype),
            self.weight,
            dims=(norm_axis, kernel_contract_axes),
        ).to(inputs.dtype)
        return output


class MlpBlock(nn.Module):
    """MLP block using DenseGeneral."""

    def __init__(self, embed_dim: int, intermediate_dim: int, compute_dtype: torch.dtype):
        super().__init__()
        self.dtype = compute_dtype

        self.wi_fused = DenseGeneral(
            in_shapes=(embed_dim,),
            out_features=(2, intermediate_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )

        self.wo = DenseGeneral(
            in_shapes=(intermediate_dim,),
            out_features=(embed_dim,),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Forward pass."""
        fused_x = self.wi_fused(x)

        gate = fused_x[..., 0, :]
        up = fused_x[..., 1, :]

        hidden = torch.mul(F.silu(gate), up).to(self.dtype)

        output = self.wo(hidden)
        return output


class RotaryEmbedding(nn.Module):
    """Rotary Position Embedding (RoPE) implementation in PyTorch."""

    def __init__(
        self,
        embedding_dims: int,
        min_timescale: float = 1.0,
        max_timescale: float = 10000.0,
        dtype: torch.dtype = torch.float32,
    ):
        super().__init__()
        if embedding_dims % 2 != 0:
            raise ValueError("Embedding dim must be even for RoPE.")
        self.embedding_dims = embedding_dims
        self.min_timescale = min_timescale
        self.max_timescale = max_timescale
        self.compute_dtype = dtype

        half_embedding_dim = embedding_dims // 2
        fraction = (2.0 * torch.arange(0, half_embedding_dim)) / embedding_dims
        timescale = (self.min_timescale * (self.max_timescale / self.min_timescale) ** fraction).to(torch.float32)
        self.register_buffer("timescale", timescale, persistent=False)

    def forward(self, inputs: torch.Tensor, position: torch.Tensor):
        """Applies RoPE."""
        position = position.unsqueeze(-1).unsqueeze(-1)
        sinusoid_inp = position / self.timescale
        sin = torch.sin(sinusoid_inp)
        cos = torch.cos(sinusoid_inp)
        first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
        first_part = first_half * cos - second_half * sin
        second_part = second_half * cos + first_half * sin
        return torch.cat(
            (first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)),
            dim=-1,
        )

    def apply_rope(self, inputs: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor):
        first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
        first_part = first_half * cos - second_half * sin
        second_part = second_half * cos + first_half * sin
        return torch.cat((first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)), dim=-1)


def custom_scaled_dot_product_attention(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attn_mask: torch.Tensor | None = None,
    scale: float = 1.0,
    is_causal: bool = False,
    num_gqa_groups: int = 1,
) -> torch.Tensor:
    """
    Custom scaled dot-product attention with GQA support for MPS compatibility.

    Args:
        query: (B, N_q, T, H) - Query tensor, N_q = num_query_heads
        key: (B, N_kv, S, H) - Key tensor, N_kv = num_kv_heads
        value: (B, N_kv, S, H) - Value tensor
        attn_mask: (B, 1, T, S) - Attention mask, optional
        scale: Scaling factor for attention scores
        is_causal: If True, apply causal masking
        num_gqa_groups: Number of query groups per KV head (N_q / N_kv)

    Returns:
        output: (B, N_q, T, H) - Attention output
    """
    B, N_q, T, H = query.shape
    _, N_kv, S, _ = key.shape

    # For GQA, repeat key and value tensors to match query heads
    if num_gqa_groups > 1:
        key = key.repeat_interleave(num_gqa_groups, dim=1)  # (B, N_q, S, H)
        value = value.repeat_interleave(num_gqa_groups, dim=1)  # (B, N_q, S, H)

    # Compute attention scores: (B, N_q, T, H) @ (B, N_q, H, S) -> (B, N_q, T, S)
    scores = torch.matmul(query, key.transpose(-1, -2)) * scale

    # Apply causal mask if needed
    if is_causal:
        causal_mask = torch.tril(torch.ones(T, S, dtype=torch.bool, device=query.device))
        scores = scores.masked_fill(~causal_mask, float("-inf"))

    # Apply attention mask if provided
    if attn_mask is not None:
        scores = scores.masked_fill(~attn_mask, float("-inf"))

    # Softmax over the last dimension (S)
    attn_weights = F.softmax(scores, dim=-1)

    # Compute output: (B, N_q, T, S) @ (B, N_q, S, H) -> (B, N_q, T, H)
    output = torch.matmul(attn_weights, value)

    return output


class CrossAttention(nn.Module):
    """Cross-Attention using DenseGeneral."""

    def __init__(
        self,
        config: EncoderConfig | DecoderConfig,
        q_embed_dim: int,
        kv_embed_dim: int,
        num_query_heads: int,
        num_kv_heads: int,
        head_dim: int,
        compute_dtype: torch.dtype,
        out_embed_dim: int | None = None,
    ):
        super().__init__()
        self.num_query_heads = num_query_heads
        self.num_kv_heads = num_kv_heads
        self.head_dim = head_dim
        self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
        self.projected_query_dim = num_query_heads * head_dim
        if num_query_heads % num_kv_heads != 0:
            raise ValueError(f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})")
        self.num_gqa_groups = num_query_heads // num_kv_heads

        # --- Projection Layers using DenseGeneral ---
        self.q_proj = DenseGeneral(
            in_shapes=(q_embed_dim,),
            out_features=(num_query_heads, head_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )
        self.k_proj = DenseGeneral(
            in_shapes=(kv_embed_dim,),
            out_features=(num_kv_heads, head_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )
        self.v_proj = DenseGeneral(
            in_shapes=(kv_embed_dim,),
            out_features=(num_kv_heads, head_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )
        self.o_proj = DenseGeneral(
            in_shapes=(num_query_heads, head_dim),
            out_features=(self.output_dim,),
            axis=(-2, -1),
            weight_dtype=compute_dtype,
        )

        # --- Rotary Embedding ---
        self.rotary_emb = RotaryEmbedding(
            embedding_dims=self.head_dim,
            max_timescale=config.rope_theta,
            dtype=compute_dtype,
        )

    def forward(
        self,
        Xq: torch.Tensor,  # (B, T, D) T = 1 in AR generation
        q_positions: torch.Tensor,  # (B, T)
        kv_positions: torch.Tensor | None = None,  # (B, S)
        attn_mask: torch.Tensor | None = None,  # None in Decoder Self Attention, Valid mask in Others
        cache: KVCache | None = None,  # None in Encoder, KVCache in Decoder
        is_causal: bool = False,
    ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
        """
        Performs attention calculation with optional KV caching.

        Args:
            Xq: Query tensor (B, T, D). T=1 during single-step decoding.
            Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
            q_positions: Positions for queries (B, T).
            kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
            attn_mask: Attention mask.
            cache: KVCache.

        Returns:
            A tuple containing:
            - output: The attention output tensor (B, T, output_dim).
            - present_kv: The K/V state to be cached for the next step ((B, N, S_new, H), (B, N, S_new, H)). For self-attn, S_new = S_past + S. For cross-attn, S_new = S_kv.
        """
        if kv_positions is None:
            kv_positions = q_positions
        original_dtype = Xq.dtype

        Xq_BxTxNxH = self.q_proj(Xq)
        Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)

        attn_k: torch.Tensor | None = cache.k if cache is not None else None
        attn_v: torch.Tensor | None = cache.v if cache is not None else None

        # Use custom attention for MPS backend, otherwise use optimized PyTorch function
        is_mps = Xq.device.type == "mps" and torch.backends.mps.is_available()
        if is_mps:
            attn_output = custom_scaled_dot_product_attention(
                query=Xq_BxNxTxH,
                key=attn_k,
                value=attn_v,
                attn_mask=attn_mask if not is_causal else None,
                scale=1.0,
                is_causal=is_causal,
                num_gqa_groups=self.num_gqa_groups,
            )
        else:
            attn_output = F.scaled_dot_product_attention(
                Xq_BxNxTxH,
                attn_k,
                attn_v,
                attn_mask=attn_mask if not is_causal else None,
                scale=1.0,
                enable_gqa=self.num_gqa_groups > 1,
                is_causal=is_causal,
            )

        attn_output = attn_output.transpose(1, 2).contiguous()  # (B, T, N, H)
        output = self.o_proj(attn_output)

        return output.to(original_dtype)


class FusedQKV(nn.Module):
    def __init__(
        self,
        in_features: int,
        out_features: int,
        bias: bool = False,
        num_q_heads: int = 1,
        q_head_dim: int = 1,
        num_kv_heads: int = 1,
        kv_head_dim: int = 1,
    ):
        super().__init__()
        self.num_q_heads = num_q_heads
        self.q_head_dim = q_head_dim
        self.num_kv_heads = num_kv_heads
        self.kv_head_dim = kv_head_dim
        self.q_output_dim = num_q_heads * q_head_dim
        self.kv_output_dim = num_kv_heads * kv_head_dim
        self.linear = nn.Linear(in_features, out_features, bias=bias)

    def forward(self, inputs: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        x = self.linear(inputs)

        q, k, v = x.split([self.q_output_dim, self.kv_output_dim, self.kv_output_dim], dim=-1)

        q = q.reshape(q.shape[:-1] + (self.num_q_heads, self.q_head_dim))
        k = k.reshape(k.shape[:-1] + (self.num_kv_heads, self.kv_head_dim))
        v = v.reshape(v.shape[:-1] + (self.num_kv_heads, self.kv_head_dim))

        return q, k, v


class SelfAttention(nn.Module):
    """Attention using DenseGeneral."""

    def __init__(
        self,
        config: EncoderConfig | DecoderConfig,
        q_embed_dim: int,
        kv_embed_dim: int,
        num_query_heads: int,
        num_kv_heads: int,
        head_dim: int,
        compute_dtype: torch.dtype,
        out_embed_dim: int | None = None,
    ):
        super().__init__()
        self.num_query_heads = num_query_heads
        self.num_kv_heads = num_kv_heads
        self.head_dim = head_dim
        self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
        self.projected_query_dim = num_query_heads * head_dim
        if num_query_heads % num_kv_heads != 0:
            raise ValueError(f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})")
        self.num_gqa_groups = num_query_heads // num_kv_heads
        self.kv_embed_dim = kv_embed_dim
        self.q_embed_dim = q_embed_dim

        # --- Projection Layers using DenseGeneral ---
        self.q_proj = DenseGeneral(
            in_shapes=(q_embed_dim,),
            out_features=(num_query_heads, head_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )
        self.k_proj = DenseGeneral(
            in_shapes=(kv_embed_dim,),
            out_features=(num_kv_heads, head_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )
        self.v_proj = DenseGeneral(
            in_shapes=(kv_embed_dim,),
            out_features=(num_kv_heads, head_dim),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )
        self.o_proj = DenseGeneral(
            in_shapes=(num_query_heads, head_dim),
            out_features=(self.output_dim,),
            axis=(-2, -1),
            weight_dtype=compute_dtype,
        )

        # --- Rotary Embedding ---
        self.rotary_emb = RotaryEmbedding(
            embedding_dims=self.head_dim,
            max_timescale=config.rope_theta,
            dtype=compute_dtype,
        )

        self.is_fused_qkv = False

    def get_linear_weight(self, dense: DenseGeneral):
        W_dg = dense.weight.data

        out_features = 1
        input_features = 1
        for dim in dense.out_features:
            out_features *= dim
        for dim in dense.in_shapes:
            input_features *= dim

        W_dg_reshaped_for_linear_T = W_dg.reshape(input_features, out_features)
        linear_weight = W_dg_reshaped_for_linear_T.transpose(0, 1).contiguous()
        return linear_weight

    def patch_fused_qkv(self):
        q_proj_weight = self.get_linear_weight(self.q_proj)
        k_proj_weight = self.get_linear_weight(self.k_proj)
        v_proj_weight = self.get_linear_weight(self.v_proj)

        self.qkv = FusedQKV(
            self.kv_embed_dim,
            (self.num_query_heads * self.head_dim + 2 * (self.num_kv_heads * self.head_dim)),
            bias=False,
            num_q_heads=self.num_query_heads,
            q_head_dim=self.head_dim,
            num_kv_heads=self.num_kv_heads,
            kv_head_dim=self.head_dim,
        )
        self.qkv.linear.weight.data = torch.cat([q_proj_weight, k_proj_weight, v_proj_weight], dim=0)

        # print(f"qkv.weight.shape: {self.qkv.linear.weight.shape}")
        self.is_fused_qkv = True

    def forward(
        self,
        X: torch.Tensor,  # (B, T, D) T = 1 in AR generation
        q_positions: torch.Tensor,  # (B, T)
        kv_positions: torch.Tensor | None = None,  # (B, S)
        attn_mask: torch.Tensor | None = None,  # None in Decoder Self Attention, Valid mask in Others
        cache: KVCache | None = None,  # None in Encoder, KVCache in Decoder
        prefill: bool = False,
        is_causal: bool = False,
        current_idx: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
        """
        Performs attention calculation with optional KV caching.
        Args:
            Xq: Query tensor (B, T, D). T=1 during single-step decoding.
            Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
            q_positions: Positions for queries (B, T).
            kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
            attn_mask: Attention mask.
            cache: KVCache.
            prefill: If True, use prefill mode.
        Returns:
            A tuple containing:
            - output: The attention output tensor (B, T, output_dim).
            - present_kv: The K/V state to be cached for the next step ((B, N, S_new, H), (B, N, S_new, H)). For self-attn, S_new = S_past + S. For cross-attn, S_new = S_kv.
        """
        if kv_positions is None:
            kv_positions = q_positions

        original_dtype = X.dtype

        if self.is_fused_qkv:
            Xq_BxTxNxH, Xk_BxSxKxH, Xv_BxSxKxH = self.qkv(X)
        else:
            Xq_BxTxNxH = self.q_proj(X)
            Xk_BxSxKxH = self.k_proj(X)
            Xv_BxSxKxH = self.v_proj(X)

        position = q_positions.unsqueeze(-1).unsqueeze(-1)
        sinusoid_inp = position / self.rotary_emb.timescale
        sin = torch.sin(sinusoid_inp)
        cos = torch.cos(sinusoid_inp)

        Xq_BxTxNxH = self.rotary_emb.apply_rope(Xq_BxTxNxH, sin, cos)
        Xk_BxSxKxH = self.rotary_emb.apply_rope(Xk_BxSxKxH, sin, cos)

        Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)

        attn_k: torch.Tensor | None = cache.k if cache is not None else None
        attn_v: torch.Tensor | None = cache.v if cache is not None else None

        Xk_BxKxSxH = Xk_BxSxKxH.transpose(1, 2)  # (B, K, S, H)
        Xv_BxKxSxH = Xv_BxSxKxH.transpose(1, 2)  # (B, K, S, H)

        if cache is None:
            attn_k = Xk_BxKxSxH
            attn_v = Xv_BxKxSxH
        elif prefill:
            attn_k, attn_v = Xk_BxKxSxH, Xv_BxKxSxH
            cache.prefill(attn_k, attn_v)
        else:
            attn_k, attn_v = cache.update(Xk_BxKxSxH, Xv_BxKxSxH, current_idx)

        # Use custom attention for MPS backend, otherwise use optimized PyTorch function
        is_mps = Xv_BxSxKxH.device.type == "mps" and torch.backends.mps.is_available()
        if is_mps:
            attn_output = custom_scaled_dot_product_attention(
                query=Xq_BxNxTxH,
                key=attn_k,
                value=attn_v,
                attn_mask=attn_mask if not is_causal else None,
                scale=1.0,
                is_causal=is_causal,
                num_gqa_groups=self.num_gqa_groups,
            )
        else:
            attn_output = F.scaled_dot_product_attention(
                Xq_BxNxTxH,
                attn_k,
                attn_v,
                attn_mask=attn_mask if not is_causal else None,
                scale=1.0,
                enable_gqa=self.num_gqa_groups > 1,
                is_causal=is_causal,
            )

        attn_output = attn_output.transpose(1, 2).contiguous()  # (B, T, N, H)
        output = self.o_proj(attn_output)

        return output.to(original_dtype)


class EncoderLayer(nn.Module):
    """Transformer Encoder Layer using DenseGeneral."""

    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
        super().__init__()
        self.config = config
        enc_config = config.encoder_config
        embed_dim = enc_config.hidden_size
        self.compute_dtype = compute_dtype

        self.pre_sa_norm = RMSNorm(
            embed_dim,
            eps=enc_config.norm_eps,
            dtype=torch.float32,
        )
        self.self_attention = SelfAttention(
            enc_config,
            q_embed_dim=embed_dim,
            kv_embed_dim=embed_dim,
            num_query_heads=enc_config.num_attention_heads,
            num_kv_heads=enc_config.num_key_value_heads,
            head_dim=enc_config.head_dim,
            compute_dtype=compute_dtype,
            out_embed_dim=embed_dim,
        )
        self.post_sa_norm = RMSNorm(
            embed_dim,
            eps=enc_config.norm_eps,
            dtype=torch.float32,
        )
        self.mlp = MlpBlock(
            embed_dim=embed_dim,
            intermediate_dim=enc_config.intermediate_size,
            compute_dtype=compute_dtype,
        )

    def forward(
        self,
        x: torch.Tensor,
        state: EncoderInferenceState,
    ) -> torch.Tensor:
        residual = x
        x_norm = self.pre_sa_norm(x).to(self.compute_dtype)

        sa_out = self.self_attention(
            X=x_norm,
            q_positions=state.positions,
            kv_positions=state.positions,
            attn_mask=state.attn_mask,
        )
        x = residual + sa_out

        residual = x
        x_norm = self.post_sa_norm(x).to(self.compute_dtype)
        mlp_out = self.mlp(x_norm)
        x = residual + mlp_out

        return x


class Encoder(nn.Module):
    """Transformer Encoder Stack using DenseGeneral."""

    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
        super().__init__()
        self.config = config
        enc_config = config.encoder_config
        self.compute_dtype = compute_dtype

        self.embedding = nn.Embedding(
            enc_config.vocab_size,
            enc_config.hidden_size,
            dtype=compute_dtype,
        )
        self.layers = nn.ModuleList([EncoderLayer(config, compute_dtype) for _ in range(enc_config.num_hidden_layers)])
        self.norm = RMSNorm(
            enc_config.hidden_size,
            eps=enc_config.norm_eps,
            dtype=torch.float32,
        )

    def forward(
        self,
        x_ids: torch.Tensor,
        state: EncoderInferenceState,
    ) -> torch.Tensor:
        x = self.embedding(x_ids)

        for layer in self.layers:
            x = layer(x, state)

        x = self.norm(x).to(self.compute_dtype)
        return x


class DecoderLayer(nn.Module):
    """Transformer Decoder Layer using DenseGeneral."""

    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
        super().__init__()
        self.config = config
        dec_config = config.decoder_config
        enc_config = config.encoder_config
        dec_embed_dim = dec_config.hidden_size
        enc_embed_dim = enc_config.hidden_size
        self.compute_dtype = compute_dtype

        # Norms
        self.pre_sa_norm = RMSNorm(
            dec_embed_dim,
            eps=dec_config.norm_eps,
            dtype=torch.float32,
        )
        self.pre_ca_norm = RMSNorm(
            dec_embed_dim,
            eps=dec_config.norm_eps,
            dtype=torch.float32,
        )
        self.pre_mlp_norm = RMSNorm(
            dec_embed_dim,
            eps=dec_config.norm_eps,
            dtype=torch.float32,
        )

        # Self-Attention (GQA) with Causal Masking
        self.self_attention = SelfAttention(
            dec_config,
            q_embed_dim=dec_embed_dim,
            kv_embed_dim=dec_embed_dim,
            num_query_heads=dec_config.num_attention_heads,
            num_kv_heads=dec_config.num_key_value_heads,
            head_dim=dec_config.head_dim,
            compute_dtype=compute_dtype,
            out_embed_dim=dec_embed_dim,
        )
        # Cross-Attention (MHA)
        self.cross_attention = CrossAttention(
            dec_config,
            q_embed_dim=dec_embed_dim,
            kv_embed_dim=enc_embed_dim,  # Note kv_embed_dim
            num_query_heads=dec_config.cross_num_attention_heads,
            num_kv_heads=dec_config.cross_num_key_value_heads,
            head_dim=dec_config.cross_head_dim,
            compute_dtype=compute_dtype,
            out_embed_dim=dec_embed_dim,
        )
        # MLP
        self.mlp = MlpBlock(
            embed_dim=dec_embed_dim,
            intermediate_dim=dec_config.intermediate_size,
            compute_dtype=compute_dtype,
        )

    def forward(
        self,
        x: torch.Tensor,
        state: DecoderInferenceState,
        self_attn_cache: KVCache | None = None,
        cross_attn_cache: KVCache | None = None,
        prefill: bool = False,
        current_idx: int = 0,
    ) -> torch.Tensor:
        residual = x
        x_norm = self.pre_sa_norm(x).to(self.compute_dtype)

        self_attn_mask = state.casual_attn_mask[None, None, current_idx]

        sa_out = self.self_attention(
            X=x_norm,  # (2, 1, D)
            q_positions=state.dec_positions,  # (2, 1)
            kv_positions=state.dec_positions,  # (2, 1)
            attn_mask=self_attn_mask,
            cache=self_attn_cache,
            prefill=prefill,
            is_causal=prefill,
            current_idx=current_idx,
        )

        x = residual + sa_out

        residual = x
        x_norm = self.pre_ca_norm(x).to(self.compute_dtype)
        ca_out = self.cross_attention(
            Xq=x_norm,
            q_positions=state.dec_positions,
            kv_positions=state.enc_positions,
            attn_mask=state.cross_attn_mask,
            cache=cross_attn_cache,
        )
        x = residual + ca_out

        residual = x
        x_norm = self.pre_mlp_norm(x).to(self.compute_dtype)
        mlp_out = self.mlp(x_norm)
        x = residual + mlp_out

        return x


class Decoder(nn.Module):
    """Transformer Decoder Stack using DenseGeneral."""

    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
        super().__init__()
        self.config = config
        dec_config = config.decoder_config
        self.num_channels = dec_config.num_channels
        self.num_layers = dec_config.num_hidden_layers

        self.embeddings = nn.ModuleList(
            [
                nn.Embedding(dec_config.vocab_size, dec_config.hidden_size, dtype=compute_dtype)
                for _ in range(self.num_channels)
            ]
        )
        self.layers = nn.ModuleList(
            [DecoderLayer(config=config, compute_dtype=compute_dtype) for _ in range(self.num_layers)]
        )

        self.norm = RMSNorm(
            dec_config.hidden_size,
            eps=dec_config.norm_eps,
            dtype=torch.float32,
        )

        self.logits_dense = DenseGeneral(
            in_shapes=(dec_config.hidden_size,),
            out_features=(self.num_channels, dec_config.vocab_size),
            axis=(-1,),
            weight_dtype=compute_dtype,
        )

    def precompute_cross_attn_cache(
        self,
        enc_out: torch.Tensor,  # (B, S, E)
    ) -> list[KVCache]:
        """
        Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
        """
        per_layer_kv_cache: list[KVCache] = []

        for layer in self.layers:
            cross_attn_module = layer.cross_attention
            k_proj = cross_attn_module.k_proj(enc_out)
            v_proj = cross_attn_module.v_proj(enc_out)

            k = k_proj.transpose(1, 2)
            v = v_proj.transpose(1, 2)

            per_layer_kv_cache.append(KVCache.from_kv(k, v))

        return per_layer_kv_cache

    def decode_step(
        self,
        tgt_ids_Bx1xC: torch.Tensor,  # [B, 1, C]
        state: DecoderInferenceState,
        current_idx: int,
    ) -> torch.Tensor:
        """
        Performs a single decoding step, managing KV caches layer by layer.
        Returns:
            A tuple containing:
            - logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
        """

        x = None
        for i in range(self.num_channels):
            channel_tokens = tgt_ids_Bx1xC[..., i]
            channel_embed = self.embeddings[i](channel_tokens)
            x = channel_embed if x is None else x + channel_embed

        for i, layer in enumerate(self.layers):
            self_cache = state.self_attn_cache[i]
            cross_cache = state.cross_attn_cache[i]
            x = layer(
                x,  # (2, 1, D)
                state,
                self_attn_cache=self_cache,
                cross_attn_cache=cross_cache,
                current_idx=current_idx,
            )

        x = self.norm(x)
        logits_Bx1xCxV = self.logits_dense(x)

        return logits_Bx1xCxV.to(torch.float32)

    def forward(self, tgt_ids_BxTxC: torch.Tensor, state: DecoderInferenceState) -> torch.Tensor:
        """
        Forward pass for the Decoder stack, managing KV caches.
        Args:
            tgt_ids_BxTxC: Target token IDs (B, T, C).
            encoder_out: Output from the encoder (B, S, E).
            tgt_positions: Positions for target sequence (B, T).
            src_positions: Positions for source sequence (B, S).
            self_attn_mask: Mask for self-attention.
            cross_attn_mask: Mask for cross-attention.
            past_key_values: List containing the self-attention KV cache for each layer
                             from the previous decoding step. `len(past_key_values)` should
                             equal `num_layers`.
            precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
                                      derived from `encoder_out`. This is passed identically
                                      to all layers.
        Returns:
            A tuple containing:
            - logits: The final output logits (B, T, C * V), cast to float32.
            - present_key_values: A list containing the updated self-attention KV cache
                                 for each layer for the *current* decoding step.
        """
        _, _, num_channels_in = tgt_ids_BxTxC.shape
        assert num_channels_in == self.num_channels, "Input channels mismatch"

        # Embeddings
        x = None
        for i in range(self.num_channels):
            channel_tokens = tgt_ids_BxTxC[..., i]
            channel_embed = self.embeddings[i](channel_tokens)
            x = channel_embed if x is None else x + channel_embed

        for i, layer in enumerate(self.layers):
            self_cache = state.self_attn_cache[i]
            cross_cache = state.cross_attn_cache[i]
            x = layer(
                x,
                state,
                self_attn_cache=self_cache,
                cross_attn_cache=cross_cache,
                prefill=True,
            )

        # Final Norm
        x = self.norm(x)
        logits_BxTxCxV = self.logits_dense(x)

        return logits_BxTxCxV.to(torch.float32)


class DiaModel(
    nn.Module,
    PyTorchModelHubMixin,
    repo_url="https://github.com/nari-labs/dia",
    pipeline_tag="text-to-speech",
    license="apache-2.0",
    coders={
        DiaConfig: (
            lambda x: x.model_dump(),
            lambda data: DiaConfig.model_validate(data),
        ),
    },
):
    """PyTorch Dia Model using DenseGeneral."""

    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
        super().__init__()
        self.config = config
        self.encoder = Encoder(config, compute_dtype)
        self.decoder = Decoder(config, compute_dtype)