diffusers 0.32.2__py3-none-any.whl → 0.33.1__py3-none-any.whl

diffusers/models/transformers/dit_transformer_2d.py

@@ -18,7 +18,7 @@ import torch.nn.functional as F
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import is_torch_version, logging
+from ...utils import logging
 from ..attention import BasicTransformerBlock
 from ..embeddings import PatchEmbed
 from ..modeling_outputs import Transformer2DModelOutput
@@ -64,7 +64,9 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
             A small constant added to the denominator in normalization layers to prevent division by zero.
     """
 
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
     _supports_gradient_checkpointing = True
+    _supports_group_offloading = False
 
     @register_to_config
     def __init__(
@@ -143,10 +145,6 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
             self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels
         )
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        if hasattr(module, "gradient_checkpointing"):
-            module.gradient_checkpointing = value
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -185,19 +183,8 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
         # 2. Blocks
         for block in self.transformer_blocks:
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-                def create_custom_forward(module, return_dict=None):
-                    def custom_forward(*inputs):
-                        if return_dict is not None:
-                            return module(*inputs, return_dict=return_dict)
-                        else:
-                            return module(*inputs)
-
-                    return custom_forward
-
-                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     None,
                     None,
@@ -205,7 +192,6 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
                     timestep,
                     cross_attention_kwargs,
                     class_labels,
-                    **ckpt_kwargs,
                 )
             else:
                 hidden_states = block(

diffusers/models/transformers/hunyuan_transformer_2d.py

@@ -244,6 +244,9 @@ class HunyuanDiT2DModel(ModelMixin, ConfigMixin):
             Whether or not to use style condition and image meta size. True for version <=1.1, False for version >= 1.2
     """
 
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm", "pooler"]
+    _supports_group_offloading = False
+
     @register_to_config
     def __init__(
         self,
@@ -277,9 +280,7 @@ class HunyuanDiT2DModel(ModelMixin, ConfigMixin):
             act_fn="silu_fp32",
         )
 
-        self.text_embedding_padding = nn.Parameter(
-            torch.randn(text_len + text_len_t5, cross_attention_dim, dtype=torch.float32)
-        )
+        self.text_embedding_padding = nn.Parameter(torch.randn(text_len + text_len_t5, cross_attention_dim))
 
         self.pos_embed = PatchEmbed(
             height=sample_size,

diffusers/models/transformers/latte_transformer_3d.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 from typing import Optional
 
 import torch
@@ -19,13 +20,14 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...models.embeddings import PixArtAlphaTextProjection, get_1d_sincos_pos_embed_from_grid
 from ..attention import BasicTransformerBlock
+from ..cache_utils import CacheMixin
 from ..embeddings import PatchEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNormSingle
 
 
-class LatteTransformer3DModel(ModelMixin, ConfigMixin):
+class LatteTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
     _supports_gradient_checkpointing = True
 
     """
@@ -65,6 +67,8 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
             The number of frames in the video-like data.
     """
 
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+
     @register_to_config
     def __init__(
         self,
@@ -162,9 +166,6 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
 
         self.gradient_checkpointing = False
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        self.gradient_checkpointing = value
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -226,20 +227,24 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
         # Prepare text embeddings for spatial block
         # batch_size num_tokens hidden_size -> (batch_size * num_frame) num_tokens hidden_size
         encoder_hidden_states = self.caption_projection(encoder_hidden_states)  # 3 120 1152
-        encoder_hidden_states_spatial = encoder_hidden_states.repeat_interleave(num_frame, dim=0).view(
-            -1, encoder_hidden_states.shape[-2], encoder_hidden_states.shape[-1]
-        )
+        encoder_hidden_states_spatial = encoder_hidden_states.repeat_interleave(
+            num_frame, dim=0, output_size=encoder_hidden_states.shape[0] * num_frame
+        ).view(-1, encoder_hidden_states.shape[-2], encoder_hidden_states.shape[-1])
 
         # Prepare timesteps for spatial and temporal block
-        timestep_spatial = timestep.repeat_interleave(num_frame, dim=0).view(-1, timestep.shape[-1])
-        timestep_temp = timestep.repeat_interleave(num_patches, dim=0).view(-1, timestep.shape[-1])
+        timestep_spatial = timestep.repeat_interleave(
+            num_frame, dim=0, output_size=timestep.shape[0] * num_frame
+        ).view(-1, timestep.shape[-1])
+        timestep_temp = timestep.repeat_interleave(
+            num_patches, dim=0, output_size=timestep.shape[0] * num_patches
+        ).view(-1, timestep.shape[-1])
 
         # Spatial and temporal transformer blocks
         for i, (spatial_block, temp_block) in enumerate(
             zip(self.transformer_blocks, self.temporal_transformer_blocks)
         ):
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-                hidden_states = torch.utils.checkpoint.checkpoint(
+                hidden_states = self._gradient_checkpointing_func(
                     spatial_block,
                     hidden_states,
                     None,  # attention_mask
@@ -248,7 +253,6 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
                     timestep_spatial,
                     None,  # cross_attention_kwargs
                     None,  # class_labels
-                    use_reentrant=False,
                 )
             else:
                 hidden_states = spatial_block(
@@ -269,10 +273,10 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
                 hidden_states = hidden_states.reshape(-1, hidden_states.shape[-2], hidden_states.shape[-1])
 
                 if i == 0 and num_frame > 1:
-                    hidden_states = hidden_states + self.temp_pos_embed
+                    hidden_states = hidden_states + self.temp_pos_embed.to(hidden_states.dtype)
 
                 if torch.is_grad_enabled() and self.gradient_checkpointing:
-                    hidden_states = torch.utils.checkpoint.checkpoint(
+                    hidden_states = self._gradient_checkpointing_func(
                         temp_block,
                         hidden_states,
                         None,  # attention_mask
@@ -281,7 +285,6 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
                         timestep_temp,
                         None,  # cross_attention_kwargs
                         None,  # class_labels
-                        use_reentrant=False,
                     )
                 else:
                     hidden_states = temp_block(
@@ -300,7 +303,9 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
                 ).permute(0, 2, 1, 3)
                 hidden_states = hidden_states.reshape(-1, hidden_states.shape[-2], hidden_states.shape[-1])
 
-        embedded_timestep = embedded_timestep.repeat_interleave(num_frame, dim=0).view(-1, embedded_timestep.shape[-1])
+        embedded_timestep = embedded_timestep.repeat_interleave(
+            num_frame, dim=0, output_size=embedded_timestep.shape[0] * num_frame
+        ).view(-1, embedded_timestep.shape[-1])
         shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
         hidden_states = self.norm_out(hidden_states)
         # Modulation

diffusers/models/transformers/lumina_nextdit2d.py

@@ -98,7 +98,7 @@ class LuminaNextDiTBlock(nn.Module):
 
         self.feed_forward = LuminaFeedForward(
             dim=dim,
-            inner_dim=4 * dim,
+            inner_dim=int(4 * 2 * dim / 3),
             multiple_of=multiple_of,
             ffn_dim_multiplier=ffn_dim_multiplier,
         )
@@ -221,6 +221,8 @@ class LuminaNextDiT2DModel(ModelMixin, ConfigMixin):
             overall scale of the model's operations.
     """
 
+    _skip_layerwise_casting_patterns = ["patch_embedder", "norm", "ffn_norm"]
+
     @register_to_config
     def __init__(
         self,

diffusers/models/transformers/pixart_transformer_2d.py

@@ -17,7 +17,7 @@ import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import is_torch_version, logging
+from ...utils import logging
 from ..attention import BasicTransformerBlock
 from ..attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
@@ -79,6 +79,7 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
 
     _supports_gradient_checkpointing = True
     _no_split_modules = ["BasicTransformerBlock", "PatchEmbed"]
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm", "adaln_single"]
 
     @register_to_config
     def __init__(
@@ -183,10 +184,6 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
                 in_features=self.config.caption_channels, hidden_size=self.inner_dim
             )
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        if hasattr(module, "gradient_checkpointing"):
-            module.gradient_checkpointing = value
-
     @property
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
     def attn_processors(self) -> Dict[str, AttentionProcessor]:
@@ -387,19 +384,8 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
         # 2. Blocks
         for block in self.transformer_blocks:
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-                def create_custom_forward(module, return_dict=None):
-                    def custom_forward(*inputs):
-                        if return_dict is not None:
-                            return module(*inputs, return_dict=return_dict)
-                        else:
-                            return module(*inputs)
-
-                    return custom_forward
-
-                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     attention_mask,
                     encoder_hidden_states,
@@ -407,7 +393,6 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
                     timestep,
                     cross_attention_kwargs,
                     None,
-                    **ckpt_kwargs,
                 )
             else:
                 hidden_states = block(

diffusers/models/transformers/prior_transformer.py

@@ -353,7 +353,11 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Pef
             attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
             attention_mask = F.pad(attention_mask, (0, self.additional_embeddings), value=0.0)
             attention_mask = (attention_mask[:, None, :] + self.causal_attention_mask).to(hidden_states.dtype)
-            attention_mask = attention_mask.repeat_interleave(self.config.num_attention_heads, dim=0)
+            attention_mask = attention_mask.repeat_interleave(
+                self.config.num_attention_heads,
+                dim=0,
+                output_size=attention_mask.shape[0] * self.config.num_attention_heads,
+            )
 
         if self.norm_in is not None:
             hidden_states = self.norm_in(hidden_states)

diffusers/models/transformers/sana_transformer.py

@@ -15,18 +15,18 @@
 from typing import Any, Dict, Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...loaders import PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ..attention_processor import (
     Attention,
     AttentionProcessor,
-    AttnProcessor2_0,
     SanaLinearAttnProcessor2_0,
 )
-from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
+from ..embeddings import PatchEmbed, PixArtAlphaTextProjection, TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNormSingle, RMSNorm
@@ -82,6 +82,109 @@ class GLUMBConv(nn.Module):
         return hidden_states
 
 
+class SanaModulatedNorm(nn.Module):
+    def __init__(self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim, elementwise_affine=elementwise_affine, eps=eps)
+
+    def forward(
+        self, hidden_states: torch.Tensor, temb: torch.Tensor, scale_shift_table: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.norm(hidden_states)
+        shift, scale = (scale_shift_table[None] + temb[:, None].to(scale_shift_table.device)).chunk(2, dim=1)
+        hidden_states = hidden_states * (1 + scale) + shift
+        return hidden_states
+
+
+class SanaCombinedTimestepGuidanceEmbeddings(nn.Module):
+    def __init__(self, embedding_dim):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.guidance_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
+
+    def forward(self, timestep: torch.Tensor, guidance: torch.Tensor = None, hidden_dtype: torch.dtype = None):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)
+
+        guidance_proj = self.guidance_condition_proj(guidance)
+        guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=hidden_dtype))
+        conditioning = timesteps_emb + guidance_emb
+
+        return self.linear(self.silu(conditioning)), conditioning
+
+
+class SanaAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("SanaAttnProcessor2_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,
+    ) -> torch.Tensor:
+        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])
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        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)
+
+        # 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)
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
 class SanaTransformerBlock(nn.Module):
     r"""
     Transformer block introduced in [Sana](https://huggingface.co/papers/2410.10629).
@@ -101,6 +204,7 @@ class SanaTransformerBlock(nn.Module):
         norm_eps: float = 1e-6,
         attention_out_bias: bool = True,
         mlp_ratio: float = 2.5,
+        qk_norm: Optional[str] = None,
     ) -> None:
         super().__init__()
 
@@ -110,6 +214,8 @@ class SanaTransformerBlock(nn.Module):
             query_dim=dim,
             heads=num_attention_heads,
             dim_head=attention_head_dim,
+            kv_heads=num_attention_heads if qk_norm is not None else None,
+            qk_norm=qk_norm,
             dropout=dropout,
             bias=attention_bias,
             cross_attention_dim=None,
@@ -121,13 +227,15 @@ class SanaTransformerBlock(nn.Module):
             self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
             self.attn2 = Attention(
                 query_dim=dim,
+                qk_norm=qk_norm,
+                kv_heads=num_cross_attention_heads if qk_norm is not None else None,
                 cross_attention_dim=cross_attention_dim,
                 heads=num_cross_attention_heads,
                 dim_head=cross_attention_head_dim,
                 dropout=dropout,
                 bias=True,
                 out_bias=attention_out_bias,
-                processor=AttnProcessor2_0(),
+                processor=SanaAttnProcessor2_0(),
             )
 
         # 3. Feed-forward
@@ -181,7 +289,7 @@ class SanaTransformerBlock(nn.Module):
         return hidden_states
 
 
-class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
     r"""
     A 2D Transformer model introduced in [Sana](https://huggingface.co/papers/2410.10629) family of models.
 
@@ -218,10 +326,15 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             Whether to use elementwise affinity in the normalization layer.
         norm_eps (`float`, defaults to `1e-6`):
             The epsilon value for the normalization layer.
+        qk_norm (`str`, *optional*, defaults to `None`):
+            The normalization to use for the query and key.
+        timestep_scale (`float`, defaults to `1.0`):
+            The scale to use for the timesteps.
     """
 
     _supports_gradient_checkpointing = True
-    _no_split_modules = ["SanaTransformerBlock", "PatchEmbed"]
+    _no_split_modules = ["SanaTransformerBlock", "PatchEmbed", "SanaModulatedNorm"]
+    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
 
     @register_to_config
     def __init__(
@@ -243,6 +356,10 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         norm_elementwise_affine: bool = False,
         norm_eps: float = 1e-6,
         interpolation_scale: Optional[int] = None,
+        guidance_embeds: bool = False,
+        guidance_embeds_scale: float = 0.1,
+        qk_norm: Optional[str] = None,
+        timestep_scale: float = 1.0,
     ) -> None:
         super().__init__()
 
@@ -250,7 +367,6 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         inner_dim = num_attention_heads * attention_head_dim
 
         # 1. Patch Embedding
-        interpolation_scale = interpolation_scale if interpolation_scale is not None else max(sample_size // 64, 1)
         self.patch_embed = PatchEmbed(
             height=sample_size,
             width=sample_size,
@@ -258,10 +374,14 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             in_channels=in_channels,
             embed_dim=inner_dim,
             interpolation_scale=interpolation_scale,
+            pos_embed_type="sincos" if interpolation_scale is not None else None,
         )
 
         # 2. Additional condition embeddings
-        self.time_embed = AdaLayerNormSingle(inner_dim)
+        if guidance_embeds:
+            self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
+        else:
+            self.time_embed = AdaLayerNormSingle(inner_dim)
 
         self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
         self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
@@ -281,6 +401,7 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
                     norm_elementwise_affine=norm_elementwise_affine,
                     norm_eps=norm_eps,
                     mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
                 )
                 for _ in range(num_layers)
             ]
@@ -288,16 +409,11 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         # 4. Output blocks
         self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
-
-        self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.norm_out = SanaModulatedNorm(inner_dim, elementwise_affine=False, eps=1e-6)
         self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
 
         self.gradient_checkpointing = False
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        if hasattr(module, "gradient_checkpointing"):
-            module.gradient_checkpointing = value
-
     @property
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
     def attn_processors(self) -> Dict[str, AttentionProcessor]:
@@ -362,7 +478,8 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
-        timestep: torch.LongTensor,
+        timestep: torch.Tensor,
+        guidance: Optional[torch.Tensor] = None,
         encoder_attention_mask: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
         attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -413,9 +530,14 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         hidden_states = self.patch_embed(hidden_states)
 
-        timestep, embedded_timestep = self.time_embed(
-            timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
-        )
+        if guidance is not None:
+            timestep, embedded_timestep = self.time_embed(
+                timestep, guidance=guidance, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            timestep, embedded_timestep = self.time_embed(
+                timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
 
         encoder_hidden_states = self.caption_projection(encoder_hidden_states)
         encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
@@ -424,21 +546,9 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         # 2. Transformer blocks
         if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-            def create_custom_forward(module, return_dict=None):
-                def custom_forward(*inputs):
-                    if return_dict is not None:
-                        return module(*inputs, return_dict=return_dict)
-                    else:
-                        return module(*inputs)
-
-                return custom_forward
-
-            ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-
             for block in self.transformer_blocks:
-                hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     attention_mask,
                     encoder_hidden_states,
@@ -446,7 +556,6 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
                     timestep,
                     post_patch_height,
                     post_patch_width,
-                    **ckpt_kwargs,
                 )
 
         else:
@@ -462,13 +571,8 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
                 )
 
         # 3. Normalization
-        shift, scale = (
-            self.scale_shift_table[None] + embedded_timestep[:, None].to(self.scale_shift_table.device)
-        ).chunk(2, dim=1)
-        hidden_states = self.norm_out(hidden_states)
+        hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table)
 
-        # 4. Modulation
-        hidden_states = hidden_states * (1 + scale) + shift
         hidden_states = self.proj_out(hidden_states)
 
         # 5. Unpatchify

diffusers/models/transformers/stable_audio_transformer.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, Optional, Union
+from typing import Dict, Optional, Union
 
 import numpy as np
 import torch
@@ -29,7 +29,7 @@ from ...models.attention_processor import (
 )
 from ...models.modeling_utils import ModelMixin
 from ...models.transformers.transformer_2d import Transformer2DModelOutput
-from ...utils import is_torch_version, logging
+from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
 
 
@@ -211,6 +211,7 @@ class StableAudioDiTModel(ModelMixin, ConfigMixin):
     """
 
     _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["preprocess_conv", "postprocess_conv", "^proj_in$", "^proj_out$", "norm"]
 
     @register_to_config
     def __init__(
@@ -345,10 +346,6 @@ class StableAudioDiTModel(ModelMixin, ConfigMixin):
         """
         self.set_attn_processor(StableAudioAttnProcessor2_0())
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        if hasattr(module, "gradient_checkpointing"):
-            module.gradient_checkpointing = value
-
     def forward(
         self,
         hidden_states: torch.FloatTensor,
@@ -415,25 +412,13 @@ class StableAudioDiTModel(ModelMixin, ConfigMixin):
 
         for block in self.transformer_blocks:
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-                def create_custom_forward(module, return_dict=None):
-                    def custom_forward(*inputs):
-                        if return_dict is not None:
-                            return module(*inputs, return_dict=return_dict)
-                        else:
-                            return module(*inputs)
-
-                    return custom_forward
-
-                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     attention_mask,
                     cross_attention_hidden_states,
                     encoder_attention_mask,
                     rotary_embedding,
-                    **ckpt_kwargs,
                 )
 
             else: