mmdit_transformer

mmdit_transformer

MMDiT (Multimodal Diffusion Transformer) implementation.

References

Esser et al. "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis" https://arxiv.org/abs/2403.03206

MMDiTTransformer

MMDiTTransformer(number_of_layers, embedding_dimension, number_of_heads, feedforward_dimension=None, dropout=0.1, attention_dropout=0.0, activation=value, normalization_type=value, positional_encoding_type=None, maximum_sequence_length=2048, maximum_decoder_length=256, timestep_embedding_dimension=256, use_query_key_norm=True, use_gating=True, bias=True, normalization_epsilon=1e-06, initializer_range=0.02)

Bases: Module

MMDiT transformer for diffusion-based action generation.

Components: - Timestep embedding network - MMDiT decoder with joint attention between observations and action tokens - Final prediction layer for action output

Shape notation

B: batch size S: observation sequence length T: action sequence length D: embedding dimension

Initialize MMDiT Transformer.

Parameters:

Name	Type	Description	Default
number_of_layers	int	Number of MMDiT layers.	required
embedding_dimension	int	Hidden dimension of the transformer.	required
number_of_heads	int	Number of attention heads.	required
feedforward_dimension	int | None	FFN hidden dimension.	None
dropout	float	Dropout rate.	0.1
attention_dropout	float	Dropout rate for attention.	0.0
activation	str	Activation function.	value
normalization_type	str	Type of normalization.	value
positional_encoding_type	str | None	Type of positional encoding.	None
maximum_sequence_length	int	Maximum observation sequence length.	2048
maximum_decoder_length	int	Maximum action sequence length.	256
timestep_embedding_dimension	int	Dimension for timestep sinusoidal embedding.	256
use_query_key_norm	bool	Whether to use QK-normalization in MMDiT layers.	True
use_gating	bool	Whether to use gating in AdaNorm (AdaLN-Zero).	True
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

Source code in src/versatil/models/layers/diffusion_transformer/mmdit_transformer.py

def __init__(
    self,
    number_of_layers: int,
    embedding_dimension: int,
    number_of_heads: int,
    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,
    positional_encoding_type: str | None = None,
    maximum_sequence_length: int = 2048,
    maximum_decoder_length: int = 256,
    timestep_embedding_dimension: int = 256,
    use_query_key_norm: bool = True,
    use_gating: bool = True,
    bias: bool = True,
    normalization_epsilon: float = 1e-6,
    initializer_range: float = 0.02,
):
    """Initialize MMDiT Transformer.

    Args:
        number_of_layers: Number of MMDiT layers.
        embedding_dimension: Hidden dimension of the transformer.
        number_of_heads: Number of attention heads.
        feedforward_dimension: FFN hidden dimension.
        dropout: Dropout rate.
        attention_dropout: Dropout rate for attention.
        activation: Activation function.
        normalization_type: Type of normalization.
        positional_encoding_type: Type of positional encoding.
        maximum_sequence_length: Maximum observation sequence length.
        maximum_decoder_length: Maximum action sequence length.
        timestep_embedding_dimension: Dimension for timestep sinusoidal embedding.
        use_query_key_norm: Whether to use QK-normalization in MMDiT layers.
        use_gating: Whether to use gating in AdaNorm (AdaLN-Zero).
        bias: Whether to use bias in linear layers.
        normalization_epsilon: Epsilon for normalization layers.
        initializer_range: Standard deviation for weight initialization.
    """
    super().__init__()
    self.embedding_dimension = embedding_dimension
    self.number_of_layers = number_of_layers
    if feedforward_dimension is None:
        feedforward_dimension = 4 * embedding_dimension
    self.timestep_embedding_network = SinusoidalPositionalEncoding1D(
        embedding_dimension=timestep_embedding_dimension,
        denominator_mode=DenominatorMode.HALF_MINUS_ONE.value,
        ordering_mode=OrderingMode.CAT_COS_SIN.value,
        position_source=PositionSource.SCALAR.value,
        precompute_encodings=False,
        temperature=10000.0,
        learnable_frequencies=False,
        mlp_activation=nn.SiLU,
        mlp_hidden_dimensions=[embedding_dimension, embedding_dimension],
    )
    self.decoder = DualStreamBidirectionalDecoder(
        number_of_layers=number_of_layers,
        embedding_dimension=embedding_dimension,
        conditioning_dimension=embedding_dimension,
        number_of_heads=number_of_heads,
        feedforward_dimension=feedforward_dimension,
        dropout=dropout,
        attention_dropout=attention_dropout,
        activation=activation,
        normalization_type=normalization_type,
        normalization_epsilon=normalization_epsilon,
        use_query_key_norm=use_query_key_norm,
        use_gating=use_gating,
        positional_encoding_type=positional_encoding_type,
        maximum_sequence_length_observation=maximum_sequence_length,
        maximum_sequence_length_action=maximum_decoder_length,
        bias=bias,
        initializer_range=initializer_range,
    )

forward

forward(decoder_hidden_states, timesteps, encoder_hidden_states, encoder_padding_mask=None, decoder_padding_mask=None)

Forward pass through the MMDiT transformer.

Parameters:

Name	Type	Description	Default
decoder_hidden_states	Tensor	Noisy action tokens (B, T, D).	required
timesteps	Tensor	Diffusion timesteps (B,).	required
encoder_hidden_states	Tensor	Observation tokens (B, S, D).	required
encoder_padding_mask	Tensor | None	Padding mask for observations (B, S).	None
decoder_padding_mask	Tensor | None	Padding mask for actions (B, T).	None

Returns:

Type	Description
Tensor	Action hidden states and conditioning, with shapes (B, T, D)
Tensor	and (B, D).

Source code in src/versatil/models/layers/diffusion_transformer/mmdit_transformer.py

def forward(
    self,
    decoder_hidden_states: torch.Tensor,
    timesteps: torch.Tensor,
    encoder_hidden_states: torch.Tensor,
    encoder_padding_mask: torch.Tensor | None = None,
    decoder_padding_mask: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Forward pass through the MMDiT transformer.

    Args:
        decoder_hidden_states: Noisy action tokens (B, T, D).
        timesteps: Diffusion timesteps (B,).
        encoder_hidden_states: Observation tokens (B, S, D).
        encoder_padding_mask: Padding mask for observations (B, S).
        decoder_padding_mask: Padding mask for actions (B, T).

    Returns:
        Action hidden states and conditioning, with shapes ``(B, T, D)``
        and ``(B, D)``.
    """
    timestep_embedding = self.timestep_embedding_network(timesteps)
    observation_output, action_output = self.decoder(
        hidden_states_observation=encoder_hidden_states,
        hidden_states_action=decoder_hidden_states,
        conditioning=timestep_embedding,
        attention_mask_observation=encoder_padding_mask,
        attention_mask_action=decoder_padding_mask,
    )
    del observation_output
    return action_output, timestep_embedding

forward_features

forward_features(decoder_hidden_states, timesteps, encoder_hidden_states, encoder_padding_mask=None, decoder_padding_mask=None)

Alias for forward kept for decoder readability.

Source code in src/versatil/models/layers/diffusion_transformer/mmdit_transformer.py

def forward_features(
    self,
    decoder_hidden_states: torch.Tensor,
    timesteps: torch.Tensor,
    encoder_hidden_states: torch.Tensor,
    encoder_padding_mask: torch.Tensor | None = None,
    decoder_padding_mask: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Alias for ``forward`` kept for decoder readability."""
    return self(
        decoder_hidden_states=decoder_hidden_states,
        timesteps=timesteps,
        encoder_hidden_states=encoder_hidden_states,
        encoder_padding_mask=encoder_padding_mask,
        decoder_padding_mask=decoder_padding_mask,
    )