bidirectional_decoder

bidirectional_decoder

Bidirectional transformer decoder for non-autoregressive generation.

BidirectionalDecoder

BidirectionalDecoder(number_of_layers, embedding_dimension, number_of_heads, number_of_key_value_heads=None, feedforward_dimension=None, dropout=0.1, attention_dropout=0.0, activation=value, normalization_type=value, attention_type=value, positional_encoding_type=None, maximum_sequence_length=2048, bias=True, normalization_epsilon=1e-06, initializer_range=0.02, use_cross_attention=True)

Bases: TransformerMixin, Module

Bidirectional transformer decoder for non-autoregressive generation.

All tokens are processed in parallel (no causal mask). Supports optional cross-attention to encoded features and conditioning cache for reusing precomputed K/V projections across repeated forward calls.

Initialize bidirectional decoder.

Parameters:

Name	Type	Description	Default
number_of_layers	int	Number of decoder layers.	required
embedding_dimension	int	Model embedding dimension.	required
number_of_heads	int	Number of attention heads.	required
number_of_key_value_heads	int | None	Number of K/V heads (for GQA).	None
feedforward_dimension	int | None	FFN hidden dimension.	None
dropout	float	Dropout probability for residual connections.	0.1
attention_dropout	float	Dropout probability for attention weights.	0.0
activation	str	Activation function (use ActivationFunction enum values).	value
normalization_type	str	Type of normalization (use NormalizationType enum values).	value
attention_type	str	Type of attention (use AttentionType enum values).	value
positional_encoding_type	str | None	Type of positional encoding (or None).	None
maximum_sequence_length	int	Maximum sequence length for positional encoding.	2048
bias	bool	Whether to use bias in linear layers.	True
normalization_epsilon	float	Epsilon for normalization layers.	1e-06
initializer_range	float	Standard deviation for weight initialization.	0.02
use_cross_attention	bool	Whether to include cross-attention blocks.	True

Source code in src/versatil/models/layers/transformer/bidirectional_decoder.py

def __init__(
    self,
    number_of_layers: int,
    embedding_dimension: int,
    number_of_heads: int,
    number_of_key_value_heads: int | None = None,
    feedforward_dimension: int | None = None,
    dropout: float = 0.1,
    attention_dropout: float = 0.0,
    activation: str = ActivationFunction.SWIGLU.value,
    normalization_type: str = NormalizationType.RMS_NORM.value,
    attention_type: str = AttentionType.GROUPED_QUERY.value,
    positional_encoding_type: str | None = None,
    maximum_sequence_length: int = 2048,
    bias: bool = True,
    normalization_epsilon: float = 1e-6,
    initializer_range: float = 0.02,
    use_cross_attention: bool = True,
):
    """Initialize bidirectional decoder.

    Args:
        number_of_layers: Number of decoder layers.
        embedding_dimension: Model embedding dimension.
        number_of_heads: Number of attention heads.
        number_of_key_value_heads: Number of K/V heads (for GQA).
        feedforward_dimension: FFN hidden dimension.
        dropout: Dropout probability for residual connections.
        attention_dropout: Dropout probability for attention weights.
        activation: Activation function (use ActivationFunction enum values).
        normalization_type: Type of normalization (use NormalizationType enum values).
        attention_type: Type of attention (use AttentionType enum values).
        positional_encoding_type: Type of positional encoding (or None).
        maximum_sequence_length: Maximum sequence length for positional encoding.
        bias: Whether to use bias in linear layers.
        normalization_epsilon: Epsilon for normalization layers.
        initializer_range: Standard deviation for weight initialization.
        use_cross_attention: Whether to include cross-attention blocks.
    """
    super().__init__()
    self.number_of_layers = number_of_layers
    self.embedding_dimension = embedding_dimension
    self.number_of_heads = number_of_heads
    self.use_cross_attention = use_cross_attention
    self.maximum_sequence_length = maximum_sequence_length
    self.initializer_range = initializer_range
    self.number_of_residual_blocks = (
        3 if use_cross_attention else 2
    )  # Self-Attention (+ Cross-Attention) + Feedforward
    self.number_of_key_value_heads, self.head_dimension = (
        self._resolve_attention_dimensions(
            embedding_dimension=embedding_dimension,
            number_of_heads=number_of_heads,
            number_of_key_value_heads=number_of_key_value_heads,
            attention_type=attention_type,
        )
    )
    self._setup_positional_encoding(
        positional_encoding_type=positional_encoding_type,
        embedding_dimension=embedding_dimension,
        maximum_sequence_length=maximum_sequence_length,
        number_of_heads=number_of_heads,
    )
    self.layers = nn.ModuleList(
        [
            TransformerDecoderLayer(
                embedding_dimension=embedding_dimension,
                number_of_heads=number_of_heads,
                number_of_key_value_heads=number_of_key_value_heads,
                feedforward_dimension=feedforward_dimension,
                dropout=dropout,
                attention_dropout=attention_dropout,
                activation=activation,
                normalization_type=normalization_type,
                attention_type=attention_type,
                use_cross_attention=use_cross_attention,
                bias=bias,
                normalization_epsilon=normalization_epsilon,
            )
            for _ in range(number_of_layers)
        ]
    )
    self.final_normalization = create_normalization_layer(
        normalization_type=normalization_type,
        dimension=embedding_dimension,
        epsilon=normalization_epsilon,
    )
    self.apply(self._init_weights)

precompute_conditioning_kv

precompute_conditioning_kv(encoded_features)

Precompute conditioning K/V for all layers for forward pass reuse.

Source code in src/versatil/models/layers/transformer/bidirectional_decoder.py

def precompute_conditioning_kv(
    self,
    encoded_features: torch.Tensor,
) -> ConditioningCache:
    """Precompute conditioning K/V for all layers for forward pass reuse."""
    return precompute_conditioning(
        layers=self.layers,
        encoded_features=encoded_features,
    )

forward

forward(hidden_states, encoded_features=None, conditioning_cache=None, query_padding_mask=None, memory_padding_mask=None)

Forward pass through bidirectional decoder.

Parameters:

Name	Type	Description	Default
hidden_states	Tensor	Query embeddings (B, query_length, D).	required
encoded_features	Tensor | None	Encoder features to cross-attend to (B, memory_length, D).	None
conditioning_cache	ConditioningCache | None	Precomputed K/V for static conditioning. When provided, encoded_features is not needed for cross-attention.	None
query_padding_mask	Tensor | None	Optional padding mask for queries (B, query_length).	None
memory_padding_mask	Tensor | None	Optional padding mask for memory (B, memory_length).	None

Returns:

Type	Description
Tensor	Output hidden states (B, query_length, D).

Raises:

Type	Description
ValueError	If use_cross_attention=True but neither encoded_features nor conditioning_cache is provided.

Source code in src/versatil/models/layers/transformer/bidirectional_decoder.py

def forward(
    self,
    hidden_states: torch.Tensor,
    encoded_features: torch.Tensor | None = None,
    conditioning_cache: ConditioningCache | None = None,
    query_padding_mask: torch.Tensor | None = None,
    memory_padding_mask: torch.Tensor | None = None,
) -> torch.Tensor:
    """Forward pass through bidirectional decoder.

    Args:
        hidden_states: Query embeddings (B, query_length, D).
        encoded_features: Encoder features to cross-attend to (B, memory_length, D).
        conditioning_cache: Precomputed K/V for static conditioning. When provided,
            encoded_features is not needed for cross-attention.
        query_padding_mask: Optional padding mask for queries (B, query_length).
        memory_padding_mask: Optional padding mask for memory (B, memory_length).

    Returns:
        Output hidden states (B, query_length, D).

    Raises:
        ValueError: If use_cross_attention=True but neither encoded_features
            nor conditioning_cache is provided.
    """
    if self.use_cross_attention and (
        encoded_features is None and conditioning_cache is None
    ):
        raise ValueError(
            "Either encoded_features or conditioning_cache must be provided "
            "when use_cross_attention=True."
        )

    query_length = hidden_states.shape[1]
    self_attention_mask = None
    if query_padding_mask is not None:
        self_attention_mask = self._expand_padding_mask(
            query_padding_mask, query_length
        )
    cross_attention_mask = None
    if memory_padding_mask is not None:
        cross_attention_mask = self._expand_padding_mask(
            memory_padding_mask, query_length
        )
    hidden_states, rope_pe = self._apply_positional_encoding(hidden_states)
    for layer_index, layer in enumerate(self.layers):
        hidden_states, _ = layer(
            hidden_states=hidden_states,
            encoded_features=encoded_features,
            self_attention_mask=self_attention_mask,
            cross_attention_mask=cross_attention_mask,
            conditioning_cache=conditioning_cache[layer_index]
            if conditioning_cache
            else None,
            positional_encoding=rope_pe,
        )
    hidden_states = self.final_normalization(hidden_states)
    return hidden_states