# AllegroTransformer3DModel

由 RhymesAI 在 [Allegro: Open the Black Box of Commercial-Level Video Generation Model](https://huggingface.co/papers/2410.15458) 中介绍了来自 [Allegro](https://github.com/rhymes-ai/Allegro) 的用于 3D 数据的 Diffusion Transformer 模型。

可以使用以下代码片段加载该模型。

```python
from diffusers import AllegroTransformer3DModel

transformer = AllegroTransformer3DModel.from_pretrained("rhymes-ai/Allegro", subfolder="transformer", dtype=torch.bfloat16).to("cuda")  # or "mps", "xpu", "cpu"
```

## AllegroTransformer3DModel[[diffusers.AllegroTransformer3DModel]]

#### diffusers.AllegroTransformer3DModel[[diffusers.AllegroTransformer3DModel]]

```python
diffusers.AllegroTransformer3DModel(patch_size: int = 2, patch_size_t: int = 1, num_attention_heads: int = 24, attention_head_dim: int = 96, in_channels: int = 4, out_channels: int = 4, num_layers: int = 32, dropout: float = 0.0, cross_attention_dim: int = 2304, attention_bias: bool = True, sample_height: int = 90, sample_width: int = 160, sample_frames: int = 22, activation_fn: str = 'gelu-approximate', norm_elementwise_affine: bool = False, norm_eps: float = 1e-06, caption_channels: int = 4096, interpolation_scale_h: float = 2.0, interpolation_scale_w: float = 2.0, interpolation_scale_t: float = 2.2)
```

[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/models/transformers/transformer_allegro.py#L174)

#### forward[[diffusers.AllegroTransformer3DModel.forward]]

```python
forward(hidden_states: Tensor, encoder_hidden_states: Tensor, timestep: LongTensor, attention_mask: typing.Optional[torch.Tensor] = None, encoder_attention_mask: typing.Optional[torch.Tensor] = None, image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None, return_dict: bool = True)
```

[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/models/transformers/transformer_allegro.py#L305)

**参数：**

hidden_states (`torch.Tensor`，形状为 `(batch_size, num_channels, num_frames, height, width)`) : 输入 `hidden_states`。

encoder_hidden_states (`torch.Tensor`，形状为 `(batch_size, sequence_len, embed_dims)`)：要使用的条件嵌入（从输入条件（例如提示词）计算得出的嵌入）。

timestep (`torch.LongTensor`)：用于指示去噪步骤。

attention_mask (`torch.Tensor`, *可选*) : 应用于 `hidden_states` 的自注意力掩码。

encoder_attention_mask (`torch.Tensor`, *可选*) : 应用于 `encoder_hidden_states` 的交叉注意力掩码。

image_rotary_emb (`tuple` of `torch.Tensor`, *可选*) : 预计算的旋转位置嵌入。

return_dict (`bool`，*可选*，默认为 `True`)：是否返回 `~models.transformer_2d.Transformer2DModelOutput` 而不是普通元组。

**返回值：**

如果 `return_dict` 为 True，则返回 `~models.transformer_2d.Transformer2DModelOutput`，否则返回一个 `tuple`，其中第一个元素是样本张量。

[AllegroTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/allegro_transformer3d#diffusers.AllegroTransformer3DModel) 的前向方法。

## Transformer2DModelOutput[[diffusers.models.modeling_outputs.Transformer2DModelOutput]]

#### diffusers.models.modeling_outputs.Transformer2DModelOutput[[diffusers.models.modeling_outputs.Transformer2DModelOutput]]

```python
diffusers.models.modeling_outputs.Transformer2DModelOutput(sample: torch.Tensor)
```

[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/models/modeling_outputs.py#L21)

**参数：**

sample (`torch.Tensor`，形状为 `(batch_size, num_channels, height, width)`；若 [Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/transformer2d#diffusers.Transformer2DModel) 为离散模式，则为 `(batch size, num_vector_embeds - 1, num_latent_pixels)`)：以 `encoder_hidden_states` 输入为条件的隐藏状态输出。若为离散模式，则返回未加噪潜在像素的概率分布。

[Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/transformer2d#diffusers.Transformer2DModel) 的输出。