FluxTransformer2DModel

A Transformer model for image-like data from Flux.

mindone.diffusers.models.transformers.transformer_flux.FluxTransformer2DModel

Bases: ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, FluxTransformer2DLoadersMixin, CacheMixin, AttentionMixin

The Transformer model introduced in Flux.

Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

PARAMETER	DESCRIPTION
patch_size	Patch size to turn the input data into small patches. TYPE: `int`, defaults to `1` DEFAULT: 1
in_channels	The number of channels in the input. TYPE: `int`, defaults to `64` DEFAULT: 64
out_channels	The number of channels in the output. If not specified, it defaults to in_channels. TYPE: `int`, *optional*, defaults to `None` DEFAULT: None
num_layers	The number of layers of dual stream DiT blocks to use. TYPE: `int`, defaults to `19` DEFAULT: 19
num_single_layers	The number of layers of single stream DiT blocks to use. TYPE: `int`, defaults to `38` DEFAULT: 38
attention_head_dim	The number of dimensions to use for each attention head. TYPE: `int`, defaults to `128` DEFAULT: 128
num_attention_heads	The number of attention heads to use. TYPE: `int`, defaults to `24` DEFAULT: 24
joint_attention_dim	The number of dimensions to use for the joint attention (embedding/channel dimension of encoder_hidden_states). TYPE: `int`, defaults to `4096` DEFAULT: 4096
pooled_projection_dim	The number of dimensions to use for the pooled projection. TYPE: `int`, defaults to `768` DEFAULT: 768
guidance_embeds	Whether to use guidance embeddings for guidance-distilled variant of the model. TYPE: `bool`, defaults to `False` DEFAULT: False
axes_dims_rope	The dimensions to use for the rotary positional embeddings. TYPE: `Tuple[int]`, defaults to `(16, 56, 56)` DEFAULT: (16, 56, 56)

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

class FluxTransformer2DModel(
    ModelMixin,
    ConfigMixin,
    PeftAdapterMixin,
    FromOriginalModelMixin,
    FluxTransformer2DLoadersMixin,
    CacheMixin,
    AttentionMixin,
):
    """
    The Transformer model introduced in Flux.

    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

    Args:
        patch_size (`int`, defaults to `1`):
            Patch size to turn the input data into small patches.
        in_channels (`int`, defaults to `64`):
            The number of channels in the input.
        out_channels (`int`, *optional*, defaults to `None`):
            The number of channels in the output. If not specified, it defaults to `in_channels`.
        num_layers (`int`, defaults to `19`):
            The number of layers of dual stream DiT blocks to use.
        num_single_layers (`int`, defaults to `38`):
            The number of layers of single stream DiT blocks to use.
        attention_head_dim (`int`, defaults to `128`):
            The number of dimensions to use for each attention head.
        num_attention_heads (`int`, defaults to `24`):
            The number of attention heads to use.
        joint_attention_dim (`int`, defaults to `4096`):
            The number of dimensions to use for the joint attention (embedding/channel dimension of
            `encoder_hidden_states`).
        pooled_projection_dim (`int`, defaults to `768`):
            The number of dimensions to use for the pooled projection.
        guidance_embeds (`bool`, defaults to `False`):
            Whether to use guidance embeddings for guidance-distilled variant of the model.
        axes_dims_rope (`Tuple[int]`, defaults to `(16, 56, 56)`):
            The dimensions to use for the rotary positional embeddings.
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]
    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
    _repeated_blocks = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]

    @register_to_config
    def __init__(
        self,
        patch_size: int = 1,
        in_channels: int = 64,
        out_channels: Optional[int] = None,
        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 = 768,
        guidance_embeds: bool = False,
        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
    ):
        super().__init__()
        self.out_channels = out_channels or in_channels
        self.inner_dim = num_attention_heads * attention_head_dim

        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)

        text_time_guidance_cls = (
            CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
        )
        self.time_text_embed = text_time_guidance_cls(
            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
        )

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

        self.transformer_blocks = nn.CellList(
            [
                FluxTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                )
                for _ in range(num_layers)
            ]
        )

        self.single_transformer_blocks = nn.CellList(
            [
                FluxSingleTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                )
                for _ in range(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,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        controlnet_block_samples=None,
        controlnet_single_block_samples=None,
        return_dict: bool = False,
        controlnet_blocks_repeat: bool = False,
    ) -> Union[ms.Tensor, Transformer2DModelOutput]:
        """
        The [`FluxTransformer2DModel`] forward method.

        Args:
            hidden_states (`ms.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`):
                Input `hidden_states`.
            encoder_hidden_states (`ms.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`):
                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.
            joint_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 `False`):
                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 joint_attention_kwargs is not None and "scale" in joint_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 {joint_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) * 1000
        if guidance is not None:
            guidance = guidance.to(hidden_states.dtype) * 1000

        temb = (
            self.time_text_embed(timestep, pooled_projections)
            if guidance is None
            else self.time_text_embed(timestep, guidance, pooled_projections)
        )
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)

        if txt_ids.ndim == 3:
            logger.warning(
                "Passing `txt_ids` 3d ms.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d mindspore Tensor"
            )
            txt_ids = txt_ids[0]
        if img_ids.ndim == 3:
            logger.warning(
                "Passing `img_ids` 3d ms.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d mindspore Tensor"
            )
            img_ids = img_ids[0]

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

        if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
            ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")
            ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds)
            joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states})

        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,
                joint_attention_kwargs=joint_attention_kwargs,
            )

            # controlnet residual
            if controlnet_block_samples is not None:
                interval_control = (len(self.transformer_blocks) + len(controlnet_block_samples) - 1) // len(
                    controlnet_block_samples
                )  # not supporting numpy
                # For Xlabs ControlNet.
                if controlnet_blocks_repeat:
                    hidden_states = (
                        hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)]
                    )
                else:
                    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,
                joint_attention_kwargs=joint_attention_kwargs,
            )

            # controlnet residual
            if controlnet_single_block_samples is not None:
                interval_control = (
                    len(self.single_transformer_blocks) + len(controlnet_single_block_samples) - 1
                ) // len(
                    controlnet_single_block_samples
                )  # not supporting numpy
                hidden_states = hidden_states + 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.models.transformers.transformer_flux.FluxTransformer2DModel.construct(hidden_states, encoder_hidden_states=None, pooled_projections=None, timestep=None, img_ids=None, txt_ids=None, guidance=None, joint_attention_kwargs=None, controlnet_block_samples=None, controlnet_single_block_samples=None, return_dict=False, controlnet_blocks_repeat=False)

The [FluxTransformer2DModel] forward method.

PARAMETER	DESCRIPTION
hidden_states	Input hidden_states. TYPE: `ms.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`
encoder_hidden_states	Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. TYPE: `ms.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)` 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.
joint_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 `False` DEFAULT: False

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

Source code in mindone/diffusers/models/transformers/transformer_flux.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,
    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
    controlnet_block_samples=None,
    controlnet_single_block_samples=None,
    return_dict: bool = False,
    controlnet_blocks_repeat: bool = False,
) -> Union[ms.Tensor, Transformer2DModelOutput]:
    """
    The [`FluxTransformer2DModel`] forward method.

    Args:
        hidden_states (`ms.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`):
            Input `hidden_states`.
        encoder_hidden_states (`ms.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`):
            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.
        joint_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 `False`):
            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 joint_attention_kwargs is not None and "scale" in joint_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 {joint_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) * 1000
    if guidance is not None:
        guidance = guidance.to(hidden_states.dtype) * 1000

    temb = (
        self.time_text_embed(timestep, pooled_projections)
        if guidance is None
        else self.time_text_embed(timestep, guidance, pooled_projections)
    )
    encoder_hidden_states = self.context_embedder(encoder_hidden_states)

    if txt_ids.ndim == 3:
        logger.warning(
            "Passing `txt_ids` 3d ms.Tensor is deprecated."
            "Please remove the batch dimension and pass it as a 2d mindspore Tensor"
        )
        txt_ids = txt_ids[0]
    if img_ids.ndim == 3:
        logger.warning(
            "Passing `img_ids` 3d ms.Tensor is deprecated."
            "Please remove the batch dimension and pass it as a 2d mindspore Tensor"
        )
        img_ids = img_ids[0]

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

    if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
        ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")
        ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds)
        joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states})

    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,
            joint_attention_kwargs=joint_attention_kwargs,
        )

        # controlnet residual
        if controlnet_block_samples is not None:
            interval_control = (len(self.transformer_blocks) + len(controlnet_block_samples) - 1) // len(
                controlnet_block_samples
            )  # not supporting numpy
            # For Xlabs ControlNet.
            if controlnet_blocks_repeat:
                hidden_states = (
                    hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)]
                )
            else:
                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,
            joint_attention_kwargs=joint_attention_kwargs,
        )

        # controlnet residual
        if controlnet_single_block_samples is not None:
            interval_control = (
                len(self.single_transformer_blocks) + len(controlnet_single_block_samples) - 1
            ) // len(
                controlnet_single_block_samples
            )  # not supporting numpy
            hidden_states = hidden_states + 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)