HunyuanVideoTransformer3DModel

A Diffusion Transformer model for 3D video-like data was introduced in HunyuanVideo: A Systematic Framework For Large Video Generative Models by Tencent.

The model can be loaded with the following code snippet.

from mindone.diffusers import HunyuanVideoTransformer3DModel
import mindspore as ms

transformer = HunyuanVideoTransformer3DModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder="transformer", mindspore_dtype=ms.bfloat16)

mindone.diffusers.HunyuanVideoTransformer3DModel

Bases: ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin

A Transformer model for video-like data used in HunyuanVideo.

PARAMETER	DESCRIPTION
in_channels	The number of channels in the input. TYPE: `int`, defaults to `16` DEFAULT: 16
out_channels	The number of channels in the output. TYPE: `int`, defaults to `16` DEFAULT: 16
num_attention_heads	The number of heads to use for multi-head attention. TYPE: `int`, defaults to `24` DEFAULT: 24
attention_head_dim	The number of channels in each head. TYPE: `int`, defaults to `128` DEFAULT: 128
num_layers	The number of layers of dual-stream blocks to use. TYPE: `int`, defaults to `20` DEFAULT: 20
num_single_layers	The number of layers of single-stream blocks to use. TYPE: `int`, defaults to `40` DEFAULT: 40
num_refiner_layers	The number of layers of refiner blocks to use. TYPE: `int`, defaults to `2` DEFAULT: 2
mlp_ratio	The ratio of the hidden layer size to the input size in the feedforward network. TYPE: `float`, defaults to `4.0` DEFAULT: 4.0
patch_size	The size of the spatial patches to use in the patch embedding layer. TYPE: `int`, defaults to `2` DEFAULT: 2
patch_size_t	The size of the tmeporal patches to use in the patch embedding layer. TYPE: `int`, defaults to `1` DEFAULT: 1
qk_norm	The normalization to use for the query and key projections in the attention layers. TYPE: `str`, defaults to `rms_norm` DEFAULT: 'rms_norm'
guidance_embeds	Whether to use guidance embeddings in the model. TYPE: `bool`, defaults to `True` DEFAULT: True
text_embed_dim	Input dimension of text embeddings from the text encoder. TYPE: `int`, defaults to `4096` DEFAULT: 4096
pooled_projection_dim	The dimension of the pooled projection of the text embeddings. TYPE: `int`, defaults to `768` DEFAULT: 768
rope_theta	The value of theta to use in the RoPE layer. TYPE: `float`, defaults to `256.0` DEFAULT: 256.0
rope_axes_dim	The dimensions of the axes to use in the RoPE layer. TYPE: `Tuple[int]`, defaults to `(16, 56, 56)` DEFAULT: (16, 56, 56)

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

class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
    r"""
    A Transformer model for video-like data used in [HunyuanVideo](https://huggingface.co/tencent/HunyuanVideo).

    Args:
        in_channels (`int`, defaults to `16`):
            The number of channels in the input.
        out_channels (`int`, defaults to `16`):
            The number of channels in the output.
        num_attention_heads (`int`, defaults to `24`):
            The number of heads to use for multi-head attention.
        attention_head_dim (`int`, defaults to `128`):
            The number of channels in each head.
        num_layers (`int`, defaults to `20`):
            The number of layers of dual-stream blocks to use.
        num_single_layers (`int`, defaults to `40`):
            The number of layers of single-stream blocks to use.
        num_refiner_layers (`int`, defaults to `2`):
            The number of layers of refiner blocks to use.
        mlp_ratio (`float`, defaults to `4.0`):
            The ratio of the hidden layer size to the input size in the feedforward network.
        patch_size (`int`, defaults to `2`):
            The size of the spatial patches to use in the patch embedding layer.
        patch_size_t (`int`, defaults to `1`):
            The size of the tmeporal patches to use in the patch embedding layer.
        qk_norm (`str`, defaults to `rms_norm`):
            The normalization to use for the query and key projections in the attention layers.
        guidance_embeds (`bool`, defaults to `True`):
            Whether to use guidance embeddings in the model.
        text_embed_dim (`int`, defaults to `4096`):
            Input dimension of text embeddings from the text encoder.
        pooled_projection_dim (`int`, defaults to `768`):
            The dimension of the pooled projection of the text embeddings.
        rope_theta (`float`, defaults to `256.0`):
            The value of theta to use in the RoPE layer.
        rope_axes_dim (`Tuple[int]`, defaults to `(16, 56, 56)`):
            The dimensions of the axes to use in the RoPE layer.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        in_channels: int = 16,
        out_channels: int = 16,
        num_attention_heads: int = 24,
        attention_head_dim: int = 128,
        num_layers: int = 20,
        num_single_layers: int = 40,
        num_refiner_layers: int = 2,
        mlp_ratio: float = 4.0,
        patch_size: int = 2,
        patch_size_t: int = 1,
        qk_norm: str = "rms_norm",
        guidance_embeds: bool = True,
        text_embed_dim: int = 4096,
        pooled_projection_dim: int = 768,
        rope_theta: float = 256.0,
        rope_axes_dim: Tuple[int] = (16, 56, 56),
    ) -> None:
        super().__init__()

        inner_dim = num_attention_heads * attention_head_dim
        out_channels = out_channels or in_channels

        # 1. Latent and condition embedders
        self.x_embedder = HunyuanVideoPatchEmbed((patch_size_t, patch_size, patch_size), in_channels, inner_dim)
        self.context_embedder = HunyuanVideoTokenRefiner(
            text_embed_dim, num_attention_heads, attention_head_dim, num_layers=num_refiner_layers
        )
        self.time_text_embed = CombinedTimestepGuidanceTextProjEmbeddings(inner_dim, pooled_projection_dim)

        # 2. RoPE
        self.rope = HunyuanVideoRotaryPosEmbed(patch_size, patch_size_t, rope_axes_dim, rope_theta)

        # 3. Dual stream transformer blocks
        self.transformer_blocks = nn.CellList(
            [
                HunyuanVideoTransformerBlock(
                    num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
                )
                for _ in range(num_layers)
            ]
        )

        # 4. Single stream transformer blocks
        self.single_transformer_blocks = nn.CellList(
            [
                HunyuanVideoSingleTransformerBlock(
                    num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
                )
                for _ in range(num_single_layers)
            ]
        )

        # 5. Output projection
        self.norm_out = AdaLayerNormContinuous(inner_dim, inner_dim, elementwise_affine=False, eps=1e-6)
        self.proj_out = nn.Dense(inner_dim, patch_size_t * patch_size * patch_size * out_channels)

        self.gradient_checkpointing = False

        self.config_patch_size = self.config.patch_size
        self.config_patch_size_t = self.config.patch_size_t

    @property
    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: nn.Cell, processors: Dict[str, AttentionProcessor]):
            if hasattr(module, "get_processor"):
                processors[f"{name}.processor"] = module.get_processor()

            for sub_name, child in module.name_cells().items():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.name_cells().items():
            fn_recursive_add_processors(name, module, processors)

        return processors

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        r"""
        Sets the attention processor to use to compute attention.

        Parameters:
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                for **all** `Attention` layers.

                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
                processor. This is strongly recommended when setting trainable attention processors.

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: nn.Cell, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.name_cells().items():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.name_cells().items():
            fn_recursive_attn_processor(name, module, processor)

    def _set_gradient_checkpointing(self, module, value=False):
        if hasattr(module, "gradient_checkpointing"):
            module.gradient_checkpointing = value

    def construct(
        self,
        hidden_states: ms.Tensor,
        timestep: ms.Tensor,
        encoder_hidden_states: ms.Tensor,
        encoder_attention_mask: ms.Tensor,
        pooled_projections: ms.Tensor,
        guidance: ms.Tensor = None,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        return_dict: bool = False,
    ) -> Union[ms.Tensor, Dict[str, ms.Tensor]]:
        if attention_kwargs is not None:
            attention_kwargs = attention_kwargs.copy()

        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__`."
            )

        batch_size, num_channels, num_frames, height, width = hidden_states.shape
        p, p_t = self.config_patch_size, self.config_patch_size_t
        post_patch_num_frames = num_frames // p_t
        post_patch_height = height // p
        post_patch_width = width // p

        # 1. RoPE
        image_rotary_emb = self.rope(hidden_states)

        # 2. Conditional embeddings
        temb = self.time_text_embed(timestep, guidance, pooled_projections)
        hidden_states = self.x_embedder(hidden_states)
        encoder_hidden_states = self.context_embedder(encoder_hidden_states, timestep, encoder_attention_mask)

        # 3. Attention mask preparation
        latent_sequence_length = hidden_states.shape[1]
        condition_sequence_length = encoder_hidden_states.shape[1]
        sequence_length = latent_sequence_length + condition_sequence_length
        attention_mask = ops.zeros((batch_size, sequence_length), dtype=ms.bool_)  # [B, N]

        effective_condition_sequence_length = encoder_attention_mask.sum(axis=1, dtype=ms.int64)  # [B,]
        effective_sequence_length = latent_sequence_length + effective_condition_sequence_length

        for i in range(batch_size):
            attention_mask[i, : effective_sequence_length[i]] = True
        attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)

        # 4. Transformer blocks
        for block in self.transformer_blocks:
            hidden_states, encoder_hidden_states = block(
                hidden_states, encoder_hidden_states, temb, attention_mask, image_rotary_emb
            )

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

        # 5. Output projection
        hidden_states = self.norm_out(hidden_states, temb)
        hidden_states = self.proj_out(hidden_states)

        hidden_states = hidden_states.reshape(
            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, -1, p_t, p, p
        )
        hidden_states = hidden_states.permute(0, 4, 1, 5, 2, 6, 3, 7)
        hidden_states = (
            hidden_states.flatten(start_dim=6, end_dim=7)
            .flatten(start_dim=4, end_dim=5)
            .flatten(start_dim=2, end_dim=3)
        )

        if not return_dict:
            return (hidden_states,)

        return Transformer2DModelOutput(sample=hidden_states)

mindone.diffusers.HunyuanVideoTransformer3DModel.attn_processors: Dict[str, AttentionProcessor] property

RETURNS	DESCRIPTION
Dict[str, AttentionProcessor]	dict of attention processors: A dictionary containing all attention processors used in the model with
Dict[str, AttentionProcessor]	indexed by its weight name.

mindone.diffusers.HunyuanVideoTransformer3DModel.set_attn_processor(processor)

Sets the attention processor to use to compute attention.

PARAMETER	DESCRIPTION
processor	The instantiated processor class or a dictionary of processor classes that will be set as the processor for all Attention layers. If processor is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainable attention processors. TYPE: `dict` of `AttentionProcessor` or only `AttentionProcessor`

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

def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
    r"""
    Sets the attention processor to use to compute attention.

    Parameters:
        processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
            The instantiated processor class or a dictionary of processor classes that will be set as the processor
            for **all** `Attention` layers.

            If `processor` is a dict, the key needs to define the path to the corresponding cross attention
            processor. This is strongly recommended when setting trainable attention processors.

    """
    count = len(self.attn_processors.keys())

    if isinstance(processor, dict) and len(processor) != count:
        raise ValueError(
            f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
            f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
        )

    def fn_recursive_attn_processor(name: str, module: nn.Cell, processor):
        if hasattr(module, "set_processor"):
            if not isinstance(processor, dict):
                module.set_processor(processor)
            else:
                module.set_processor(processor.pop(f"{name}.processor"))

        for sub_name, child in module.name_cells().items():
            fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

    for name, module in self.name_cells().items():
        fn_recursive_attn_processor(name, module, processor)

mindone.diffusers.models.modeling_outputs.Transformer2DModelOutput dataclass

Bases: BaseOutput

The output of [Transformer2DModel].

PARAMETER	DESCRIPTION
`(batch	The hidden states output conditioned on the encoder_hidden_states input. If discrete, returns probability distributions for the unnoised latent pixels. TYPE: size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete

Source code in mindone/diffusers/models/modeling_outputs.py

@dataclass
class Transformer2DModelOutput(BaseOutput):
    """
    The output of [`Transformer2DModel`].

    Args:
        sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` or
        `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
            distributions for the unnoised latent pixels.
    """

    sample: "ms.Tensor"  # noqa: F821