diffusers 0.15.1__py3-none-any.whl → 0.16.1__py3-none-any.whl

diffusers/models/attention_processor.py

@@ -149,6 +149,9 @@ class Attention(nn.Module):
         is_lora = hasattr(self, "processor") and isinstance(
             self.processor, (LoRAAttnProcessor, LoRAXFormersAttnProcessor)
         )
+        is_custom_diffusion = hasattr(self, "processor") and isinstance(
+            self.processor, (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor)
+        )
 
         if use_memory_efficient_attention_xformers:
             if self.added_kv_proj_dim is not None:
@@ -192,6 +195,17 @@ class Attention(nn.Module):
                 )
                 processor.load_state_dict(self.processor.state_dict())
                 processor.to(self.processor.to_q_lora.up.weight.device)
+            elif is_custom_diffusion:
+                processor = CustomDiffusionXFormersAttnProcessor(
+                    train_kv=self.processor.train_kv,
+                    train_q_out=self.processor.train_q_out,
+                    hidden_size=self.processor.hidden_size,
+                    cross_attention_dim=self.processor.cross_attention_dim,
+                    attention_op=attention_op,
+                )
+                processor.load_state_dict(self.processor.state_dict())
+                if hasattr(self.processor, "to_k_custom_diffusion"):
+                    processor.to(self.processor.to_k_custom_diffusion.weight.device)
             else:
                 processor = XFormersAttnProcessor(attention_op=attention_op)
         else:
@@ -203,6 +217,16 @@ class Attention(nn.Module):
                 )
                 processor.load_state_dict(self.processor.state_dict())
                 processor.to(self.processor.to_q_lora.up.weight.device)
+            elif is_custom_diffusion:
+                processor = CustomDiffusionAttnProcessor(
+                    train_kv=self.processor.train_kv,
+                    train_q_out=self.processor.train_q_out,
+                    hidden_size=self.processor.hidden_size,
+                    cross_attention_dim=self.processor.cross_attention_dim,
+                )
+                processor.load_state_dict(self.processor.state_dict())
+                if hasattr(self.processor, "to_k_custom_diffusion"):
+                    processor.to(self.processor.to_k_custom_diffusion.weight.device)
             else:
                 processor = AttnProcessor()
 
@@ -459,6 +483,84 @@ class LoRAAttnProcessor(nn.Module):
         return hidden_states
 
 
+class CustomDiffusionAttnProcessor(nn.Module):
+    def __init__(
+        self,
+        train_kv=True,
+        train_q_out=True,
+        hidden_size=None,
+        cross_attention_dim=None,
+        out_bias=True,
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.train_kv = train_kv
+        self.train_q_out = train_q_out
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+
+        # `_custom_diffusion` id for easy serialization and loading.
+        if self.train_kv:
+            self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        if self.train_q_out:
+            self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+            self.to_out_custom_diffusion = nn.ModuleList([])
+            self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+            self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+    def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        batch_size, sequence_length, _ = hidden_states.shape
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+        if self.train_q_out:
+            query = self.to_q_custom_diffusion(hidden_states)
+        else:
+            query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            crossattn = False
+            encoder_hidden_states = hidden_states
+        else:
+            crossattn = True
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        if self.train_kv:
+            key = self.to_k_custom_diffusion(encoder_hidden_states)
+            value = self.to_v_custom_diffusion(encoder_hidden_states)
+        else:
+            key = attn.to_k(encoder_hidden_states)
+            value = attn.to_v(encoder_hidden_states)
+
+        if crossattn:
+            detach = torch.ones_like(key)
+            detach[:, :1, :] = detach[:, :1, :] * 0.0
+            key = detach * key + (1 - detach) * key.detach()
+            value = detach * value + (1 - detach) * value.detach()
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if self.train_q_out:
+            # linear proj
+            hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+            # dropout
+            hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+        else:
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+
+        return hidden_states
+
+
 class AttnAddedKVProcessor:
     def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
         residual = hidden_states
@@ -699,6 +801,91 @@ class LoRAXFormersAttnProcessor(nn.Module):
         return hidden_states
 
 
+class CustomDiffusionXFormersAttnProcessor(nn.Module):
+    def __init__(
+        self,
+        train_kv=True,
+        train_q_out=False,
+        hidden_size=None,
+        cross_attention_dim=None,
+        out_bias=True,
+        dropout=0.0,
+        attention_op: Optional[Callable] = None,
+    ):
+        super().__init__()
+        self.train_kv = train_kv
+        self.train_q_out = train_q_out
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.attention_op = attention_op
+
+        # `_custom_diffusion` id for easy serialization and loading.
+        if self.train_kv:
+            self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        if self.train_q_out:
+            self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+            self.to_out_custom_diffusion = nn.ModuleList([])
+            self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+            self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+    def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if self.train_q_out:
+            query = self.to_q_custom_diffusion(hidden_states)
+        else:
+            query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            crossattn = False
+            encoder_hidden_states = hidden_states
+        else:
+            crossattn = True
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        if self.train_kv:
+            key = self.to_k_custom_diffusion(encoder_hidden_states)
+            value = self.to_v_custom_diffusion(encoder_hidden_states)
+        else:
+            key = attn.to_k(encoder_hidden_states)
+            value = attn.to_v(encoder_hidden_states)
+
+        if crossattn:
+            detach = torch.ones_like(key)
+            detach[:, :1, :] = detach[:, :1, :] * 0.0
+            key = detach * key + (1 - detach) * key.detach()
+            value = detach * value + (1 - detach) * value.detach()
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        hidden_states = xformers.ops.memory_efficient_attention(
+            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
+        )
+        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if self.train_q_out:
+            # linear proj
+            hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+            # dropout
+            hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+        else:
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
 class SlicedAttnProcessor:
     def __init__(self, slice_size):
         self.slice_size = slice_size
@@ -834,4 +1021,6 @@ AttentionProcessor = Union[
     AttnAddedKVProcessor2_0,
     LoRAAttnProcessor,
     LoRAXFormersAttnProcessor,
+    CustomDiffusionAttnProcessor,
+    CustomDiffusionXFormersAttnProcessor,
 ]

diffusers/models/controlnet.py

@@ -457,6 +457,7 @@ class ControlNetModel(ModelMixin, ConfigMixin):
         timestep_cond: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guess_mode: bool = False,
         return_dict: bool = True,
     ) -> Union[ControlNetOutput, Tuple]:
         # check channel order
@@ -557,8 +558,14 @@ class ControlNetModel(ModelMixin, ConfigMixin):
         mid_block_res_sample = self.controlnet_mid_block(sample)
 
         # 6. scaling
-        down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
-        mid_block_res_sample *= conditioning_scale
+        if guess_mode:
+            scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1)  # 0.1 to 1.0
+            scales *= conditioning_scale
+            down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
+            mid_block_res_sample *= scales[-1]  # last one
+        else:
+            down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
+            mid_block_res_sample *= conditioning_scale
 
         if self.config.global_pool_conditions:
             down_block_res_samples = [

diffusers/models/embeddings.py

@@ -377,3 +377,69 @@ class CombinedTimestepLabelEmbeddings(nn.Module):
         conditioning = timesteps_emb + class_labels  # (N, D)
 
         return conditioning
+
+
+class TextTimeEmbedding(nn.Module):
+    def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(encoder_dim)
+        self.pool = AttentionPooling(num_heads, encoder_dim)
+        self.proj = nn.Linear(encoder_dim, time_embed_dim)
+        self.norm2 = nn.LayerNorm(time_embed_dim)
+
+    def forward(self, hidden_states):
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.pool(hidden_states)
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+class AttentionPooling(nn.Module):
+    # Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54
+
+    def __init__(self, num_heads, embed_dim, dtype=None):
+        super().__init__()
+        self.dtype = dtype
+        self.positional_embedding = nn.Parameter(torch.randn(1, embed_dim) / embed_dim**0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
+        self.num_heads = num_heads
+        self.dim_per_head = embed_dim // self.num_heads
+
+    def forward(self, x):
+        bs, length, width = x.size()
+
+        def shape(x):
+            # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+            x = x.view(bs, -1, self.num_heads, self.dim_per_head)
+            # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+            x = x.transpose(1, 2)
+            # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+            x = x.reshape(bs * self.num_heads, -1, self.dim_per_head)
+            # (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length)
+            x = x.transpose(1, 2)
+            return x
+
+        class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(x.dtype)
+        x = torch.cat([class_token, x], dim=1)  # (bs, length+1, width)
+
+        # (bs*n_heads, class_token_length, dim_per_head)
+        q = shape(self.q_proj(class_token))
+        # (bs*n_heads, length+class_token_length, dim_per_head)
+        k = shape(self.k_proj(x))
+        v = shape(self.v_proj(x))
+
+        # (bs*n_heads, class_token_length, length+class_token_length):
+        scale = 1 / math.sqrt(math.sqrt(self.dim_per_head))
+        weight = torch.einsum("bct,bcs->bts", q * scale, k * scale)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+
+        # (bs*n_heads, dim_per_head, class_token_length)
+        a = torch.einsum("bts,bcs->bct", weight, v)
+
+        # (bs, length+1, width)
+        a = a.reshape(bs, -1, 1).transpose(1, 2)
+
+        return a[:, 0, :]  # cls_token

diffusers/models/modeling_pytorch_flax_utils.py

@@ -110,6 +110,12 @@ def load_flax_weights_in_pytorch_model(pt_model, flax_state):
                     .replace("_1", ".1")
                     .replace("_2", ".2")
                     .replace("_3", ".3")
+                    .replace("_4", ".4")
+                    .replace("_5", ".5")
+                    .replace("_6", ".6")
+                    .replace("_7", ".7")
+                    .replace("_8", ".8")
+                    .replace("_9", ".9")
                 )
 
         flax_key = ".".join(flax_key_tuple_array)

diffusers/models/modeling_utils.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 import inspect
+import itertools
 import os
 from functools import partial
 from typing import Any, Callable, List, Optional, Tuple, Union
@@ -60,7 +61,8 @@ if is_safetensors_available():
 
 def get_parameter_device(parameter: torch.nn.Module):
     try:
-        return next(parameter.parameters()).device
+        parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
+        return next(parameters_and_buffers).device
     except StopIteration:
         # For torch.nn.DataParallel compatibility in PyTorch 1.5
 
@@ -75,7 +77,8 @@ def get_parameter_device(parameter: torch.nn.Module):
 
 def get_parameter_dtype(parameter: torch.nn.Module):
     try:
-        return next(parameter.parameters()).dtype
+        parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
+        return next(parameters_and_buffers).dtype
     except StopIteration:
         # For torch.nn.DataParallel compatibility in PyTorch 1.5
 

diffusers/models/transformer_2d.py

@@ -225,7 +225,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
             hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
                 When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
                 hidden_states
-            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
                 Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                 self-attention.
             timestep ( `torch.long`, *optional*):

diffusers/models/unet_2d_condition.py

@@ -23,7 +23,7 @@ from ..configuration_utils import ConfigMixin, register_to_config
 from ..loaders import UNet2DConditionLoadersMixin
 from ..utils import BaseOutput, logging
 from .attention_processor import AttentionProcessor, AttnProcessor
-from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
+from .embeddings import GaussianFourierProjection, TextTimeEmbedding, TimestepEmbedding, Timesteps
 from .modeling_utils import ModelMixin
 from .unet_2d_blocks import (
     CrossAttnDownBlock2D,
@@ -97,11 +97,16 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         class_embed_type (`str`, *optional*, defaults to None):
             The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
             `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to None):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
         num_class_embeds (`int`, *optional*, defaults to None):
             Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
             class conditioning with `class_embed_type` equal to `None`.
         time_embedding_type (`str`, *optional*, default to `positional`):
             The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, default to `None`):
+            An optional override for the dimension of the projected time embedding.
         time_embedding_act_fn (`str`, *optional*, default to `None`):
             Optional activation function to use on the time embeddings only one time before they as passed to the rest
             of the unet. Choose from `silu`, `mish`, `gelu`, and `swish`.
@@ -155,12 +160,14 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         dual_cross_attention: bool = False,
         use_linear_projection: bool = False,
         class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
         num_class_embeds: Optional[int] = None,
         upcast_attention: bool = False,
         resnet_time_scale_shift: str = "default",
         resnet_skip_time_act: bool = False,
         resnet_out_scale_factor: int = 1.0,
         time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
         time_embedding_act_fn: Optional[str] = None,
         timestep_post_act: Optional[str] = None,
         time_cond_proj_dim: Optional[int] = None,
@@ -170,6 +177,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         class_embeddings_concat: bool = False,
         mid_block_only_cross_attention: Optional[bool] = None,
         cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
     ):
         super().__init__()
 
@@ -214,7 +222,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
 
         # time
         if time_embedding_type == "fourier":
-            time_embed_dim = block_out_channels[0] * 2
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
             if time_embed_dim % 2 != 0:
                 raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
             self.time_proj = GaussianFourierProjection(
@@ -222,7 +230,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             )
             timestep_input_dim = time_embed_dim
         elif time_embedding_type == "positional":
-            time_embed_dim = block_out_channels[0] * 4
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
 
             self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
             timestep_input_dim = block_out_channels[0]
@@ -248,7 +256,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         if class_embed_type is None and num_class_embeds is not None:
             self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
         elif class_embed_type == "timestep":
-            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
         elif class_embed_type == "identity":
             self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
         elif class_embed_type == "projection":
@@ -273,6 +281,18 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         else:
             self.class_embedding = None
 
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None or 'text'.")
+
         if time_embedding_act_fn is None:
             self.time_embed_act = None
         elif time_embedding_act_fn == "swish":
@@ -437,7 +457,18 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             self.conv_norm_out = nn.GroupNorm(
                 num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
             )
-            self.conv_act = nn.SiLU()
+
+            if act_fn == "swish":
+                self.conv_act = lambda x: F.silu(x)
+            elif act_fn == "mish":
+                self.conv_act = nn.Mish()
+            elif act_fn == "silu":
+                self.conv_act = nn.SiLU()
+            elif act_fn == "gelu":
+                self.conv_act = nn.GELU()
+            else:
+                raise ValueError(f"Unsupported activation function: {act_fn}")
+
         else:
             self.conv_norm_out = None
             self.conv_act = None
@@ -648,7 +679,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
 
         t_emb = self.time_proj(timesteps)
 
-        # timesteps does not contain any weights and will always return f32 tensors
+        # `Timesteps` does not contain any weights and will always return f32 tensors
         # but time_embedding might actually be running in fp16. so we need to cast here.
         # there might be better ways to encapsulate this.
         t_emb = t_emb.to(dtype=self.dtype)
@@ -662,6 +693,10 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             if self.config.class_embed_type == "timestep":
                 class_labels = self.time_proj(class_labels)
 
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
             class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
 
             if self.config.class_embeddings_concat:
@@ -669,6 +704,10 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             else:
                 emb = emb + class_emb
 
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+            emb = emb + aug_emb
+
         if self.time_embed_act is not None:
             emb = self.time_embed_act(emb)
 

diffusers/models/vae.py

@@ -212,6 +212,7 @@ class Decoder(nn.Module):
         sample = z
         sample = self.conv_in(sample)
 
+        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
         if self.training and self.gradient_checkpointing:
 
             def create_custom_forward(module):
@@ -222,6 +223,7 @@ class Decoder(nn.Module):
 
             # middle
             sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
+            sample = sample.to(upscale_dtype)
 
             # up
             for up_block in self.up_blocks:
@@ -229,6 +231,7 @@ class Decoder(nn.Module):
         else:
             # middle
             sample = self.mid_block(sample)
+            sample = sample.to(upscale_dtype)
 
             # up
             for up_block in self.up_blocks:

diffusers/pipelines/__init__.py

@@ -44,6 +44,14 @@ except OptionalDependencyNotAvailable:
 else:
     from .alt_diffusion import AltDiffusionImg2ImgPipeline, AltDiffusionPipeline
     from .audioldm import AudioLDMPipeline
+    from .deepfloyd_if import (
+        IFImg2ImgPipeline,
+        IFImg2ImgSuperResolutionPipeline,
+        IFInpaintingPipeline,
+        IFInpaintingSuperResolutionPipeline,
+        IFPipeline,
+        IFSuperResolutionPipeline,
+    )
     from .latent_diffusion import LDMTextToImagePipeline
     from .paint_by_example import PaintByExamplePipeline
     from .semantic_stable_diffusion import SemanticStableDiffusionPipeline

diffusers/pipelines/alt_diffusion/modeling_roberta_series.py

@@ -56,7 +56,7 @@ class RobertaSeriesConfig(XLMRobertaConfig):
 
 
 class RobertaSeriesModelWithTransformation(RobertaPreTrainedModel):
-    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"logit_scale"]
     _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
     base_model_prefix = "roberta"
     config_class = RobertaSeriesConfig
@@ -65,6 +65,10 @@ class RobertaSeriesModelWithTransformation(RobertaPreTrainedModel):
         super().__init__(config)
         self.roberta = XLMRobertaModel(config)
         self.transformation = nn.Linear(config.hidden_size, config.project_dim)
+        self.has_pre_transformation = getattr(config, "has_pre_transformation", False)
+        if self.has_pre_transformation:
+            self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim)
+            self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.post_init()
 
     def forward(
@@ -95,15 +99,26 @@ class RobertaSeriesModelWithTransformation(RobertaPreTrainedModel):
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_attention_mask,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
+            output_hidden_states=True if self.has_pre_transformation else output_hidden_states,
             return_dict=return_dict,
         )
 
-        projection_state = self.transformation(outputs.last_hidden_state)
-
-        return TransformationModelOutput(
-            projection_state=projection_state,
-            last_hidden_state=outputs.last_hidden_state,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
-        )
+        if self.has_pre_transformation:
+            sequence_output2 = outputs["hidden_states"][-2]
+            sequence_output2 = self.pre_LN(sequence_output2)
+            projection_state2 = self.transformation_pre(sequence_output2)
+
+            return TransformationModelOutput(
+                projection_state=projection_state2,
+                last_hidden_state=outputs.last_hidden_state,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+            )
+        else:
+            projection_state = self.transformation(outputs.last_hidden_state)
+            return TransformationModelOutput(
+                projection_state=projection_state,
+                last_hidden_state=outputs.last_hidden_state,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+            )

diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -57,6 +57,14 @@ class AltDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
     library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
 
+    In addition the pipeline inherits the following loading methods:
+        - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
+        - *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
+        - *Ckpt*: [`loaders.FromCkptMixin.from_ckpt`]
+
+    as well as the following saving methods:
+        - *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.

diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -96,6 +96,14 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
     library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
 
+    In addition the pipeline inherits the following loading methods:
+        - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
+        - *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
+        - *Ckpt*: [`loaders.FromCkptMixin.from_ckpt`]
+
+    as well as the following saving methods:
+        - *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.

diffusers/pipelines/audioldm/pipeline_audioldm.py

@@ -293,7 +293,7 @@ class AudioLDMPipeline(DiffusionPipeline):
 
         waveform = self.vocoder(mel_spectrogram)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
-        waveform = waveform.cpu()
+        waveform = waveform.cpu().float()
         return waveform
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs