[core] Allegro T2V

src/diffusers/__init__.py

@@ -77,9 +77,11 @@
 else:
     _import_structure["models"].extend(
         [
+            "AllegroTransformer3DModel",
             "AsymmetricAutoencoderKL",
             "AuraFlowTransformer2DModel",
             "AutoencoderKL",
+            "AutoencoderKLAllegro",
             "AutoencoderKLCogVideoX",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderOobleck",
@@ -237,6 +239,7 @@
 else:
     _import_structure["pipelines"].extend(
         [
+            "AllegroPipeline",
             "AltDiffusionImg2ImgPipeline",
             "AltDiffusionPipeline",
             "AmusedImg2ImgPipeline",
@@ -556,9 +559,11 @@
         from .utils.dummy_pt_objects import *  # noqa F403
     else:
         from .models import (
+            AllegroTransformer3DModel,
             AsymmetricAutoencoderKL,
             AuraFlowTransformer2DModel,
             AutoencoderKL,
+            AutoencoderKLAllegro,
             AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
@@ -697,6 +702,7 @@
         from .utils.dummy_torch_and_transformers_objects import *  # noqa F403
     else:
         from .pipelines import (
+            AllegroPipeline,
             AltDiffusionImg2ImgPipeline,
             AltDiffusionPipeline,
             AmusedImg2ImgPipeline,

src/diffusers/models/__init__.py

@@ -28,6 +28,7 @@
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
     _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
     _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_allegro"] = ["AutoencoderKLAllegro"]
     _import_structure["autoencoders.autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
@@ -54,6 +55,7 @@
     _import_structure["transformers.stable_audio_transformer"] = ["StableAudioDiTModel"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
+    _import_structure["transformers.transformer_allegro"] = ["AllegroTransformer3DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
@@ -81,6 +83,7 @@
         from .autoencoders import (
             AsymmetricAutoencoderKL,
             AutoencoderKL,
+            AutoencoderKLAllegro,
             AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
@@ -97,6 +100,7 @@
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .transformers import (
+            AllegroTransformer3DModel,
             AuraFlowTransformer2DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,

src/diffusers/models/attention_processor.py

@@ -1521,6 +1521,100 @@ def __call__(
         return hidden_states, encoder_hidden_states
 
 
+class AllegroAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
+    used in the Allegro model. It applies a s normalization layer and rotary embedding on query and key vector.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AllegroAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None and not attn.is_cross_attention:
+            from .embeddings import apply_rotary_emb_allegro
+
+            query = apply_rotary_emb_allegro(query, image_rotary_emb[0], image_rotary_emb[1])
+            key = apply_rotary_emb_allegro(key, image_rotary_emb[0], image_rotary_emb[1])
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
 class AuraFlowAttnProcessor2_0:
     """Attention processor used typically in processing Aura Flow."""
 

src/diffusers/models/autoencoders/__init__.py

@@ -1,5 +1,6 @@
 from .autoencoder_asym_kl import AsymmetricAutoencoderKL
 from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_allegro import AutoencoderKLAllegro
 from .autoencoder_kl_cogvideox import AutoencoderKLCogVideoX
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_oobleck import AutoencoderOobleck