BriaTransformer2DModel

A modified flux Transformer model from Bria

mindone.diffusers.BriaTransformer2DModel

Bases: ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin

The Transformer model introduced in Flux. Based on FluxPipeline with several changes: - no pooled embeddings - We use zero padding for prompts - No guidance embedding since this is not a distilled version Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

PARAMETER	DESCRIPTION
patch_size	Patch size to turn the input data into small patches. TYPE: `int` DEFAULT: 1
in_channels	The number of channels in the input. TYPE: `int`, *optional*, defaults to 16 DEFAULT: 64
num_layers	The number of layers of MMDiT blocks to use. TYPE: `int`, *optional*, defaults to 18 DEFAULT: 19
num_single_layers	The number of layers of single DiT blocks to use. TYPE: `int`, *optional*, defaults to 18 DEFAULT: 38
attention_head_dim	The number of channels in each head. TYPE: `int`, *optional*, defaults to 64 DEFAULT: 128
num_attention_heads	The number of heads to use for multi-head attention. TYPE: `int`, *optional*, defaults to 18 DEFAULT: 24
joint_attention_dim	The number of encoder_hidden_states dimensions to use. TYPE: `int`, *optional* DEFAULT: 4096
pooled_projection_dim	Number of dimensions to use when projecting the pooled_projections. TYPE: `int` DEFAULT: None
guidance_embeds	Whether to use guidance embeddings. TYPE: `bool`, defaults to False DEFAULT: False

Source code in mindone/diffusers/models/transformers/transformer_bria.py

class BriaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
    """
    The Transformer model introduced in Flux. Based on FluxPipeline with several changes:
    - no pooled embeddings
    - We use zero padding for prompts
    - No guidance embedding since this is not a distilled version
    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

    Parameters:
        patch_size (`int`): Patch size to turn the input data into small patches.
        in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
        num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
        num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
        attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
        num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
        joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
        pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
        guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        patch_size: int = 1,
        in_channels: int = 64,
        num_layers: int = 19,
        num_single_layers: int = 38,
        attention_head_dim: int = 128,
        num_attention_heads: int = 24,
        joint_attention_dim: int = 4096,
        pooled_projection_dim: int = None,
        guidance_embeds: bool = False,
        axes_dims_rope: List[int] = [16, 56, 56],
        rope_theta=10000,
        time_theta=10000,
    ):
        super().__init__()
        self.out_channels = in_channels
        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim

        self.pos_embed = BriaEmbedND(theta=rope_theta, axes_dim=axes_dims_rope)

        self.time_embed = BriaTimestepProjEmbeddings(embedding_dim=self.inner_dim, time_theta=time_theta)
        if guidance_embeds:
            self.guidance_embed = BriaTimestepProjEmbeddings(embedding_dim=self.inner_dim)

        self.context_embedder = mint.nn.Linear(self.config.joint_attention_dim, self.inner_dim)
        self.x_embedder = mint.nn.Linear(self.config.in_channels, self.inner_dim)

        self.transformer_blocks = nn.CellList(
            [
                BriaTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=self.config.num_attention_heads,
                    attention_head_dim=self.config.attention_head_dim,
                )
                for i in range(self.config.num_layers)
            ]
        )

        self.single_transformer_blocks = nn.CellList(
            [
                BriaSingleTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=self.config.num_attention_heads,
                    attention_head_dim=self.config.attention_head_dim,
                )
                for i in range(self.config.num_single_layers)
            ]
        )

        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
        self.proj_out = mint.nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)

        self.gradient_checkpointing = False

    def construct(
        self,
        hidden_states: ms.Tensor,
        encoder_hidden_states: ms.Tensor = None,
        pooled_projections: ms.Tensor = None,
        timestep: ms.Tensor = None,
        img_ids: ms.Tensor = None,
        txt_ids: ms.Tensor = None,
        guidance: ms.Tensor = None,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        return_dict: bool = True,
        controlnet_block_samples=None,
        controlnet_single_block_samples=None,
    ) -> Union[Tuple[ms.Tensor], Transformer2DModelOutput]:
        """
        The [`BriaTransformer2DModel`] forward method.

        Args:
            hidden_states (`ms.Tensor` of shape `(batch size, channel, height, width)`):
                Input `hidden_states`.
            encoder_hidden_states (`ms.Tensor` of shape `(batch size, sequence_len, embed_dims)`):
                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
            pooled_projections (`ms.Tensor` of shape `(batch_size, projection_dim)`): Embeddings projected
                from the embeddings of input conditions.
            timestep ( `ms.Tensor`):
                Used to indicate denoising step.
            block_controlnet_hidden_states: (`list` of `ms.Tensor`):
                A list of tensors that if specified are added to the residuals of transformer blocks.
            attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
                tuple.

        Returns:
            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
            `tuple` where the first element is the sample tensor.
        """
        if attention_kwargs is not None and "scale" in attention_kwargs:
            # weight the lora layers by setting `lora_scale` for each PEFT layer here
            # and remove `lora_scale` from each PEFT layer at the end.
            # scale_lora_layers & unscale_lora_layers maybe contains some operation forbidden in graph mode
            raise RuntimeError(
                f"You are trying to set scaling of lora layer by passing {attention_kwargs['scale']=}. "
                f"However it's not allowed in on-the-fly model forwarding. "
                f"Please manually call `scale_lora_layers(model, lora_scale)` before model forwarding and "
                f"`unscale_lora_layers(model, lora_scale)` after model forwarding. "
                f"For example, it can be done in a pipeline call like `StableDiffusionPipeline.__call__`."
            )
        hidden_states = self.x_embedder(hidden_states)

        timestep = timestep.to(hidden_states.dtype)
        if guidance is not None:
            guidance = guidance.to(hidden_states.dtype)
        else:
            guidance = None

        temb = self.time_embed(timestep, dtype=hidden_states.dtype)

        if guidance:
            temb += self.guidance_embed(guidance, dtype=hidden_states.dtype)

        encoder_hidden_states = self.context_embedder(encoder_hidden_states)

        if len(txt_ids.shape) == 3:
            txt_ids = txt_ids[0]

        if len(img_ids.shape) == 3:
            img_ids = img_ids[0]

        ids = mint.cat((txt_ids, img_ids), dim=0)
        image_rotary_emb = self.pos_embed(ids)

        for index_block, block in enumerate(self.transformer_blocks):
            encoder_hidden_states, hidden_states = block(
                hidden_states=hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                temb=temb,
                image_rotary_emb=image_rotary_emb,
            )

            # controlnet residual
            if controlnet_block_samples is not None:
                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
                interval_control = int(np.ceil(interval_control))
                hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]

        for index_block, block in enumerate(self.single_transformer_blocks):
            encoder_hidden_states, hidden_states = block(
                hidden_states=hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                temb=temb,
                image_rotary_emb=image_rotary_emb,
            )

            # controlnet residual
            if controlnet_single_block_samples is not None:
                interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
                interval_control = int(np.ceil(interval_control))
                hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
                    hidden_states[:, encoder_hidden_states.shape[1] :, ...]
                    + controlnet_single_block_samples[index_block // interval_control]
                )

        hidden_states = self.norm_out(hidden_states, temb)
        output = self.proj_out(hidden_states)

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)

mindone.diffusers.BriaTransformer2DModel.construct(hidden_states, encoder_hidden_states=None, pooled_projections=None, timestep=None, img_ids=None, txt_ids=None, guidance=None, attention_kwargs=None, return_dict=True, controlnet_block_samples=None, controlnet_single_block_samples=None)

The [BriaTransformer2DModel] forward method.

PARAMETER	DESCRIPTION
hidden_states	Input hidden_states. TYPE: `ms.Tensor` of shape `(batch size, channel, height, width)`
encoder_hidden_states	Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. TYPE: `ms.Tensor` of shape `(batch size, sequence_len, embed_dims)` DEFAULT: None
pooled_projections	Embeddings projected from the embeddings of input conditions. TYPE: `ms.Tensor` of shape `(batch_size, projection_dim)` DEFAULT: None
timestep	Used to indicate denoising step. TYPE: `ms.Tensor` DEFAULT: None
block_controlnet_hidden_states	(list of ms.Tensor): A list of tensors that if specified are added to the residuals of transformer blocks.
attention_kwargs	A kwargs dictionary that if specified is passed along to the AttentionProcessor as defined under self.processor in diffusers.models.attention_processor. TYPE: `dict`, *optional* DEFAULT: None
return_dict	Whether or not to return a [~models.transformer_2d.Transformer2DModelOutput] instead of a plain tuple. TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

RETURNS	DESCRIPTION
Union[Tuple[Tensor], Transformer2DModelOutput]	If return_dict is True, an [~models.transformer_2d.Transformer2DModelOutput] is returned, otherwise a
Union[Tuple[Tensor], Transformer2DModelOutput]	tuple where the first element is the sample tensor.

Source code in mindone/diffusers/models/transformers/transformer_bria.py

def construct(
    self,
    hidden_states: ms.Tensor,
    encoder_hidden_states: ms.Tensor = None,
    pooled_projections: ms.Tensor = None,
    timestep: ms.Tensor = None,
    img_ids: ms.Tensor = None,
    txt_ids: ms.Tensor = None,
    guidance: ms.Tensor = None,
    attention_kwargs: Optional[Dict[str, Any]] = None,
    return_dict: bool = True,
    controlnet_block_samples=None,
    controlnet_single_block_samples=None,
) -> Union[Tuple[ms.Tensor], Transformer2DModelOutput]:
    """
    The [`BriaTransformer2DModel`] forward method.

    Args:
        hidden_states (`ms.Tensor` of shape `(batch size, channel, height, width)`):
            Input `hidden_states`.
        encoder_hidden_states (`ms.Tensor` of shape `(batch size, sequence_len, embed_dims)`):
            Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
        pooled_projections (`ms.Tensor` of shape `(batch_size, projection_dim)`): Embeddings projected
            from the embeddings of input conditions.
        timestep ( `ms.Tensor`):
            Used to indicate denoising step.
        block_controlnet_hidden_states: (`list` of `ms.Tensor`):
            A list of tensors that if specified are added to the residuals of transformer blocks.
        attention_kwargs (`dict`, *optional*):
            A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
            `self.processor` in
            [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
        return_dict (`bool`, *optional*, defaults to `True`):
            Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
            tuple.

    Returns:
        If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
        `tuple` where the first element is the sample tensor.
    """
    if attention_kwargs is not None and "scale" in attention_kwargs:
        # weight the lora layers by setting `lora_scale` for each PEFT layer here
        # and remove `lora_scale` from each PEFT layer at the end.
        # scale_lora_layers & unscale_lora_layers maybe contains some operation forbidden in graph mode
        raise RuntimeError(
            f"You are trying to set scaling of lora layer by passing {attention_kwargs['scale']=}. "
            f"However it's not allowed in on-the-fly model forwarding. "
            f"Please manually call `scale_lora_layers(model, lora_scale)` before model forwarding and "
            f"`unscale_lora_layers(model, lora_scale)` after model forwarding. "
            f"For example, it can be done in a pipeline call like `StableDiffusionPipeline.__call__`."
        )
    hidden_states = self.x_embedder(hidden_states)

    timestep = timestep.to(hidden_states.dtype)
    if guidance is not None:
        guidance = guidance.to(hidden_states.dtype)
    else:
        guidance = None

    temb = self.time_embed(timestep, dtype=hidden_states.dtype)

    if guidance:
        temb += self.guidance_embed(guidance, dtype=hidden_states.dtype)

    encoder_hidden_states = self.context_embedder(encoder_hidden_states)

    if len(txt_ids.shape) == 3:
        txt_ids = txt_ids[0]

    if len(img_ids.shape) == 3:
        img_ids = img_ids[0]

    ids = mint.cat((txt_ids, img_ids), dim=0)
    image_rotary_emb = self.pos_embed(ids)

    for index_block, block in enumerate(self.transformer_blocks):
        encoder_hidden_states, hidden_states = block(
            hidden_states=hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            temb=temb,
            image_rotary_emb=image_rotary_emb,
        )

        # controlnet residual
        if controlnet_block_samples is not None:
            interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
            interval_control = int(np.ceil(interval_control))
            hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]

    for index_block, block in enumerate(self.single_transformer_blocks):
        encoder_hidden_states, hidden_states = block(
            hidden_states=hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            temb=temb,
            image_rotary_emb=image_rotary_emb,
        )

        # controlnet residual
        if controlnet_single_block_samples is not None:
            interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
            interval_control = int(np.ceil(interval_control))
            hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
                hidden_states[:, encoder_hidden_states.shape[1] :, ...]
                + controlnet_single_block_samples[index_block // interval_control]
            )

    hidden_states = self.norm_out(hidden_states, temb)
    output = self.proj_out(hidden_states)

    if not return_dict:
        return (output,)

    return Transformer2DModelOutput(sample=output)