mps: fix torch.where in svd codepath.

* Fix possible re-conversion issues after extracting from diffusers

Properly rename specific vae keys.

* Whoops

benchmarks/base_classes.py

@@ -162,6 +162,25 @@ class LCMLoRATextToImageBenchmark(TextToImageBenchmark):
             guidance_scale=1.0,
         )
 
+    def benchmark(self, args):
+        flush()
+
+        print(f"[INFO] {self.pipe.__class__.__name__}: Running benchmark with: {vars(args)}\n")
+
+        time = benchmark_fn(self.run_inference, self.pipe, args)  # in seconds.
+        memory = bytes_to_giga_bytes(torch.cuda.max_memory_allocated())  # in GBs.
+        benchmark_info = BenchmarkInfo(time=time, memory=memory)
+
+        pipeline_class_name = str(self.pipe.__class__.__name__)
+        flush()
+        csv_dict = generate_csv_dict(
+            pipeline_cls=pipeline_class_name, ckpt=self.lora_id, args=args, benchmark_info=benchmark_info
+        )
+        filepath = self.get_result_filepath(args)
+        write_to_csv(filepath, csv_dict)
+        print(f"Logs written to: {filepath}")
+        flush()
+
 
 class ImageToImageBenchmark(TextToImageBenchmark):
     pipeline_class = AutoPipelineForImage2Image

docs/source/en/api/models/asymmetricautoencoderkl.md

@@ -49,12 +49,12 @@ make_image_grid([original_image, mask_image, image], rows=1, cols=3)
 
 ## AsymmetricAutoencoderKL
 
-[[autodoc]] models.autoencoder_asym_kl.AsymmetricAutoencoderKL
+[[autodoc]] models.autoencoders.autoencoder_asym_kl.AsymmetricAutoencoderKL
 
 ## AutoencoderKLOutput
 
-[[autodoc]] models.autoencoder_kl.AutoencoderKLOutput
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
 
 ## DecoderOutput
 
-[[autodoc]] models.vae.DecoderOutput
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/autoencoder_tiny.md

@@ -54,4 +54,4 @@ image
 
 ## AutoencoderTinyOutput
 
-[[autodoc]] models.autoencoder_tiny.AutoencoderTinyOutput
+[[autodoc]] models.autoencoders.autoencoder_tiny.AutoencoderTinyOutput

docs/source/en/api/models/autoencoderkl.md

@@ -36,11 +36,11 @@ model = AutoencoderKL.from_single_file(url)
 
 ## AutoencoderKLOutput
 
-[[autodoc]] models.autoencoder_kl.AutoencoderKLOutput
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
 
 ## DecoderOutput
 
-[[autodoc]] models.vae.DecoderOutput
+[[autodoc]] models.autoencoders.vae.DecoderOutput
 
 ## FlaxAutoencoderKL
 

docs/source/en/training/unconditional_training.md

@@ -186,7 +186,7 @@ accelerate launch train_unconditional.py \
 If you're training with more than one GPU, add the `--multi_gpu` parameter to the training command:
 
 ```bash
-accelerate launch --mixed_precision="fp16" --multi_gpu train_unconditional.py \
+accelerate launch --multi_gpu train_unconditional.py \
   --dataset_name="huggan/flowers-102-categories" \
   --output_dir="ddpm-ema-flowers-64" \
   --mixed_precision="fp16" \

examples/dreambooth/train_dreambooth_lora.py

@@ -64,39 +64,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__)
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(
     repo_id: str,
     images=None,

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -64,39 +64,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__)
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(
     repo_id: str,
     images=None,

examples/text_to_image/README.md

@@ -101,8 +101,8 @@ accelerate launch --mixed_precision="fp16" train_text_to_image.py \
 
 Once the training is finished the model will be saved in the `output_dir` specified in the command. In this example it's `sd-pokemon-model`. To load the fine-tuned model for inference just pass that path to `StableDiffusionPipeline`
 
-
 ```python
+import torch
 from diffusers import StableDiffusionPipeline
 
 model_path = "path_to_saved_model"
@@ -114,12 +114,13 @@ image.save("yoda-pokemon.png")
 ```
 
 Checkpoints only save the unet, so to run inference from a checkpoint, just load the unet
+
 ```python
+import torch
 from diffusers import StableDiffusionPipeline, UNet2DConditionModel
 
 model_path = "path_to_saved_model"
-
-unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-<N>/unet")
+unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-<N>/unet", torch_dtype=torch.float16)
 
 pipe = StableDiffusionPipeline.from_pretrained("<initial model>", unet=unet, torch_dtype=torch.float16)
 pipe.to("cuda")

examples/text_to_image/train_text_to_image_lora.py

@@ -54,39 +54,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__, log_level="INFO")
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None):
     img_str = ""
     for i, image in enumerate(images):

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -63,39 +63,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__)
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(
     repo_id: str,
     images=None,

scripts/convert_consistency_decoder.py

@@ -12,9 +12,9 @@ from safetensors.torch import load_file as stl
 from tqdm import tqdm
 
 from diffusers import AutoencoderKL, ConsistencyDecoderVAE, DiffusionPipeline, StableDiffusionPipeline, UNet2DModel
+from diffusers.models.autoencoders.vae import Encoder
 from diffusers.models.embeddings import TimestepEmbedding
 from diffusers.models.unet_2d_blocks import ResnetDownsampleBlock2D, ResnetUpsampleBlock2D, UNetMidBlock2D
-from diffusers.models.vae import Encoder
 
 
 args = ArgumentParser()

scripts/convert_diffusers_to_original_stable_diffusion.py

@@ -159,6 +159,14 @@ vae_conversion_map_attn = [
     ("proj_out.", "proj_attn."),
 ]
 
+# This is probably not the most ideal solution, but it does work.
+vae_extra_conversion_map = [
+    ("to_q", "q"),
+    ("to_k", "k"),
+    ("to_v", "v"),
+    ("to_out.0", "proj_out"),
+]
+
 
 def reshape_weight_for_sd(w):
     # convert HF linear weights to SD conv2d weights
@@ -178,11 +186,20 @@ def convert_vae_state_dict(vae_state_dict):
             mapping[k] = v
     new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
     weights_to_convert = ["q", "k", "v", "proj_out"]
+    keys_to_rename = {}
     for k, v in new_state_dict.items():
         for weight_name in weights_to_convert:
             if f"mid.attn_1.{weight_name}.weight" in k:
                 print(f"Reshaping {k} for SD format")
                 new_state_dict[k] = reshape_weight_for_sd(v)
+        for weight_name, real_weight_name in vae_extra_conversion_map:
+            if f"mid.attn_1.{weight_name}.weight" in k or f"mid.attn_1.{weight_name}.bias" in k:
+                keys_to_rename[k] = k.replace(weight_name, real_weight_name)
+    for k, v in keys_to_rename.items():
+        if k in new_state_dict:
+            print(f"Renaming {k} to {v}")
+            new_state_dict[v] = reshape_weight_for_sd(new_state_dict[k])
+            del new_state_dict[k]
     return new_state_dict
 
 

src/diffusers/loaders/single_file.py

@@ -169,10 +169,12 @@ class FromSingleFileMixin:
         load_safety_checker = kwargs.pop("load_safety_checker", True)
         prediction_type = kwargs.pop("prediction_type", None)
         text_encoder = kwargs.pop("text_encoder", None)
+        text_encoder_2 = kwargs.pop("text_encoder_2", None)
         vae = kwargs.pop("vae", None)
         controlnet = kwargs.pop("controlnet", None)
         adapter = kwargs.pop("adapter", None)
         tokenizer = kwargs.pop("tokenizer", None)
+        tokenizer_2 = kwargs.pop("tokenizer_2", None)
 
         torch_dtype = kwargs.pop("torch_dtype", None)
 
@@ -274,8 +276,10 @@ class FromSingleFileMixin:
             load_safety_checker=load_safety_checker,
             prediction_type=prediction_type,
             text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
             vae=vae,
             tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
             original_config_file=original_config_file,
             config_files=config_files,
             local_files_only=local_files_only,

src/diffusers/models/__init__.py

@@ -26,11 +26,11 @@ _import_structure = {}
 
 if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
-    _import_structure["autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
-    _import_structure["autoencoder_kl"] = ["AutoencoderKL"]
-    _import_structure["autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
-    _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
-    _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
+    _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
     _import_structure["controlnetxs"] = ["ControlNetXSModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
@@ -58,11 +58,13 @@ if is_flax_available():
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     if is_torch_available():
         from .adapter import MultiAdapter, T2IAdapter
-        from .autoencoder_asym_kl import AsymmetricAutoencoderKL
-        from .autoencoder_kl import AutoencoderKL
-        from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
-        from .autoencoder_tiny import AutoencoderTiny
-        from .consistency_decoder_vae import ConsistencyDecoderVAE
+        from .autoencoders import (
+            AsymmetricAutoencoderKL,
+            AutoencoderKL,
+            AutoencoderKLTemporalDecoder,
+            AutoencoderTiny,
+            ConsistencyDecoderVAE,
+        )
         from .controlnet import ControlNetModel
         from .controlnetxs import ControlNetXSModel
         from .dual_transformer_2d import DualTransformer2DModel

src/diffusers/models/autoencoders/__init__.py

@@ -0,0 +1,5 @@
+from .autoencoder_asym_kl import AsymmetricAutoencoderKL
+from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
+from .autoencoder_tiny import AutoencoderTiny
+from .consistency_decoder_vae import ConsistencyDecoderVAE

src/diffusers/models/autoencoders/autoencoder_asym_kl.py

@@ -16,10 +16,10 @@ from typing import Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils.accelerate_utils import apply_forward_hook
-from .modeling_outputs import AutoencoderKLOutput
-from .modeling_utils import ModelMixin
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
 from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder, MaskConditionDecoder
 
 

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -16,10 +16,10 @@ from typing import Dict, Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..loaders import FromOriginalVAEMixin
-from ..utils.accelerate_utils import apply_forward_hook
-from .attention_processor import (
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalVAEMixin
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
     Attention,
@@ -27,8 +27,8 @@ from .attention_processor import (
     AttnAddedKVProcessor,
     AttnProcessor,
 )
-from .modeling_outputs import AutoencoderKLOutput
-from .modeling_utils import ModelMixin
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
 from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
 
 

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -16,14 +16,14 @@ from typing import Dict, Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..loaders import FromOriginalVAEMixin
-from ..utils import is_torch_version
-from ..utils.accelerate_utils import apply_forward_hook
-from .attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
-from .modeling_outputs import AutoencoderKLOutput
-from .modeling_utils import ModelMixin
-from .unet_3d_blocks import MidBlockTemporalDecoder, UpBlockTemporalDecoder
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalVAEMixin
+from ...utils import is_torch_version
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from ..unet_3d_blocks import MidBlockTemporalDecoder, UpBlockTemporalDecoder
 from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder
 
 

src/diffusers/models/autoencoders/autoencoder_tiny.py

@@ -18,10 +18,10 @@ from typing import Optional, Tuple, Union
 
 import torch
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput
-from ..utils.accelerate_utils import apply_forward_hook
-from .modeling_utils import ModelMixin
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_utils import ModelMixin
 from .vae import DecoderOutput, DecoderTiny, EncoderTiny
 
 

src/diffusers/models/autoencoders/consistency_decoder_vae.py

@@ -18,20 +18,20 @@ import torch
 import torch.nn.functional as F
 from torch import nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..schedulers import ConsistencyDecoderScheduler
-from ..utils import BaseOutput
-from ..utils.accelerate_utils import apply_forward_hook
-from ..utils.torch_utils import randn_tensor
-from .attention_processor import (
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...schedulers import ConsistencyDecoderScheduler
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ...utils.torch_utils import randn_tensor
+from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
-from .modeling_utils import ModelMixin
-from .unet_2d import UNet2DModel
+from ..modeling_utils import ModelMixin
+from ..unet_2d import UNet2DModel
 from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder
 
 
@@ -153,7 +153,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         self.use_slicing = False
         self.use_tiling = False
 
-    # Copied from diffusers.models.autoencoder_kl.AutoencoderKL.enable_tiling
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_tiling
     def enable_tiling(self, use_tiling: bool = True):
         r"""
         Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
@@ -162,7 +162,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         """
         self.use_tiling = use_tiling
 
-    # Copied from diffusers.models.autoencoder_kl.AutoencoderKL.disable_tiling
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.disable_tiling
     def disable_tiling(self):
         r"""
         Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
@@ -170,7 +170,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         """
         self.enable_tiling(False)
 
-    # Copied from diffusers.models.autoencoder_kl.AutoencoderKL.enable_slicing
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_slicing
     def enable_slicing(self):
         r"""
         Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
@@ -178,7 +178,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         """
         self.use_slicing = True
 
-    # Copied from diffusers.models.autoencoder_kl.AutoencoderKL.disable_slicing
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.disable_slicing
     def disable_slicing(self):
         r"""
         Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
@@ -333,14 +333,14 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
 
         return DecoderOutput(sample=x_0)
 
-    # Copied from diffusers.models.autoencoder_kl.AutoencoderKL.blend_v
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.blend_v
     def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
         blend_extent = min(a.shape[2], b.shape[2], blend_extent)
         for y in range(blend_extent):
             b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
         return b
 
-    # Copied from diffusers.models.autoencoder_kl.AutoencoderKL.blend_h
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.blend_h
     def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
         blend_extent = min(a.shape[3], b.shape[3], blend_extent)
         for x in range(blend_extent):

src/diffusers/models/autoencoders/vae.py

@@ -18,11 +18,11 @@ import numpy as np
 import torch
 import torch.nn as nn
 
-from ..utils import BaseOutput, is_torch_version
-from ..utils.torch_utils import randn_tensor
-from .activations import get_activation
-from .attention_processor import SpatialNorm
-from .unet_2d_blocks import (
+from ...utils import BaseOutput, is_torch_version
+from ...utils.torch_utils import randn_tensor
+from ..activations import get_activation
+from ..attention_processor import SpatialNorm
+from ..unet_2d_blocks import (
     AutoencoderTinyBlock,
     UNetMidBlock2D,
     get_down_block,

src/diffusers/models/controlnetxs.py

@@ -26,7 +26,7 @@ from ..utils import BaseOutput, logging
 from .attention_processor import (
     AttentionProcessor,
 )
-from .autoencoder_kl import AutoencoderKL
+from .autoencoders import AutoencoderKL
 from .lora import LoRACompatibleConv
 from .modeling_utils import ModelMixin
 from .unet_2d_blocks import (

src/diffusers/models/resnet.py

@@ -1334,7 +1334,7 @@ class AlphaBlender(nn.Module):
 
             alpha = torch.where(
                 image_only_indicator.bool(),
-                torch.ones(1, 1, device=image_only_indicator.device),
+                torch.ones(1, 1, device=image_only_indicator.device, dtype=self.mix_factor.dtype),
                 torch.sigmoid(self.mix_factor)[..., None],
             )
 

src/diffusers/models/vq_model.py

@@ -20,8 +20,8 @@ import torch.nn as nn
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..utils import BaseOutput
 from ..utils.accelerate_utils import apply_forward_hook
+from .autoencoders.vae import Decoder, DecoderOutput, Encoder, VectorQuantizer
 from .modeling_utils import ModelMixin
-from .vae import Decoder, DecoderOutput, Encoder, VectorQuantizer
 
 
 @dataclass

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -23,6 +23,7 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -655,6 +656,65 @@ class AltDiffusionPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """
         See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -25,6 +25,7 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -715,6 +716,65 @@ class AltDiffusionImg2ImgPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """
         See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -1153,7 +1153,9 @@ def download_from_original_stable_diffusion_ckpt(
     vae_path=None,
     vae=None,
     text_encoder=None,
+    text_encoder_2=None,
     tokenizer=None,
+    tokenizer_2=None,
     config_files=None,
 ) -> DiffusionPipeline:
     """
@@ -1232,7 +1234,9 @@ def download_from_original_stable_diffusion_ckpt(
         StableDiffusionInpaintPipeline,
         StableDiffusionPipeline,
         StableDiffusionUpscalePipeline,
+        StableDiffusionXLControlNetInpaintPipeline,
         StableDiffusionXLImg2ImgPipeline,
+        StableDiffusionXLInpaintPipeline,
         StableDiffusionXLPipeline,
         StableUnCLIPImg2ImgPipeline,
         StableUnCLIPPipeline,
@@ -1339,7 +1343,11 @@ def download_from_original_stable_diffusion_ckpt(
         else:
             pipeline_class = StableDiffusionXLPipeline if model_type == "SDXL" else StableDiffusionXLImg2ImgPipeline
 
-    if num_in_channels is None and pipeline_class == StableDiffusionInpaintPipeline:
+    if num_in_channels is None and pipeline_class in [
+        StableDiffusionInpaintPipeline,
+        StableDiffusionXLInpaintPipeline,
+        StableDiffusionXLControlNetInpaintPipeline,
+    ]:
         num_in_channels = 9
     if num_in_channels is None and pipeline_class == StableDiffusionUpscalePipeline:
         num_in_channels = 7
@@ -1686,7 +1694,9 @@ def download_from_original_stable_diffusion_ckpt(
                 feature_extractor=feature_extractor,
             )
     elif model_type in ["SDXL", "SDXL-Refiner"]:
-        if model_type == "SDXL":
+        is_refiner = model_type == "SDXL-Refiner"
+
+        if (is_refiner is False) and (tokenizer is None):
             try:
                 tokenizer = CLIPTokenizer.from_pretrained(
                     "openai/clip-vit-large-patch14", local_files_only=local_files_only
@@ -1695,7 +1705,11 @@ def download_from_original_stable_diffusion_ckpt(
                 raise ValueError(
                     f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'."
                 )
+
+        if (is_refiner is False) and (text_encoder is None):
             text_encoder = convert_ldm_clip_checkpoint(checkpoint, local_files_only=local_files_only)
+
+        if tokenizer_2 is None:
             try:
                 tokenizer_2 = CLIPTokenizer.from_pretrained(
                     "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!", local_files_only=local_files_only
@@ -1705,95 +1719,69 @@ def download_from_original_stable_diffusion_ckpt(
                     f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'laion/CLIP-ViT-bigG-14-laion2B-39B-b160k' with `pad_token` set to '!'."
                 )
 
+        if text_encoder_2 is None:
             config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
             config_kwargs = {"projection_dim": 1280}
+            prefix = "conditioner.embedders.0.model." if is_refiner else "conditioner.embedders.1.model."
+
             text_encoder_2 = convert_open_clip_checkpoint(
                 checkpoint,
                 config_name,
-                prefix="conditioner.embedders.1.model.",
+                prefix=prefix,
                 has_projection=True,
                 local_files_only=local_files_only,
                 **config_kwargs,
             )
 
-            if is_accelerate_available():  # SBM Now move model to cpu.
-                if model_type in ["SDXL", "SDXL-Refiner"]:
-                    for param_name, param in converted_unet_checkpoint.items():
-                        set_module_tensor_to_device(unet, param_name, "cpu", value=param)
+        if is_accelerate_available():  # SBM Now move model to cpu.
+            for param_name, param in converted_unet_checkpoint.items():
+                set_module_tensor_to_device(unet, param_name, "cpu", value=param)
 
-            if controlnet:
-                pipe = pipeline_class(
-                    vae=vae,
-                    text_encoder=text_encoder,
-                    tokenizer=tokenizer,
-                    text_encoder_2=text_encoder_2,
-                    tokenizer_2=tokenizer_2,
-                    unet=unet,
-                    controlnet=controlnet,
-                    scheduler=scheduler,
-                    force_zeros_for_empty_prompt=True,
-                )
-            elif adapter:
-                pipe = pipeline_class(
-                    vae=vae,
-                    text_encoder=text_encoder,
-                    tokenizer=tokenizer,
-                    text_encoder_2=text_encoder_2,
-                    tokenizer_2=tokenizer_2,
-                    unet=unet,
-                    adapter=adapter,
-                    scheduler=scheduler,
-                    force_zeros_for_empty_prompt=True,
-                )
-            else:
-                pipe = pipeline_class(
-                    vae=vae,
-                    text_encoder=text_encoder,
-                    tokenizer=tokenizer,
-                    text_encoder_2=text_encoder_2,
-                    tokenizer_2=tokenizer_2,
-                    unet=unet,
-                    scheduler=scheduler,
-                    force_zeros_for_empty_prompt=True,
-                )
-        else:
-            tokenizer = None
-            text_encoder = None
-            try:
-                tokenizer_2 = CLIPTokenizer.from_pretrained(
-                    "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!", local_files_only=local_files_only
-                )
-            except Exception:
-                raise ValueError(
-                    f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'laion/CLIP-ViT-bigG-14-laion2B-39B-b160k' with `pad_token` set to '!'."
-                )
-            config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
-            config_kwargs = {"projection_dim": 1280}
-            text_encoder_2 = convert_open_clip_checkpoint(
-                checkpoint,
-                config_name,
-                prefix="conditioner.embedders.0.model.",
-                has_projection=True,
-                local_files_only=local_files_only,
-                **config_kwargs,
-            )
-
-            if is_accelerate_available():  # SBM Now move model to cpu.
-                if model_type in ["SDXL", "SDXL-Refiner"]:
-                    for param_name, param in converted_unet_checkpoint.items():
-                        set_module_tensor_to_device(unet, param_name, "cpu", value=param)
-
-            pipe = StableDiffusionXLImg2ImgPipeline(
+        if controlnet:
+            pipe = pipeline_class(
                 vae=vae,
                 text_encoder=text_encoder,
                 tokenizer=tokenizer,
                 text_encoder_2=text_encoder_2,
                 tokenizer_2=tokenizer_2,
                 unet=unet,
+                controlnet=controlnet,
                 scheduler=scheduler,
-                requires_aesthetics_score=True,
-                force_zeros_for_empty_prompt=False,
+                force_zeros_for_empty_prompt=True,
             )
+        elif adapter:
+            pipe = pipeline_class(
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                text_encoder_2=text_encoder_2,
+                tokenizer_2=tokenizer_2,
+                unet=unet,
+                adapter=adapter,
+                scheduler=scheduler,
+                force_zeros_for_empty_prompt=True,
+            )
+
+        else:
+            pipeline_kwargs = {
+                "vae": vae,
+                "text_encoder": text_encoder,
+                "tokenizer": tokenizer,
+                "text_encoder_2": text_encoder_2,
+                "tokenizer_2": tokenizer_2,
+                "unet": unet,
+                "scheduler": scheduler,
+            }
+
+            if (pipeline_class == StableDiffusionXLImg2ImgPipeline) or (
+                pipeline_class == StableDiffusionXLInpaintPipeline
+            ):
+                pipeline_kwargs.update({"requires_aesthetics_score": is_refiner})
+
+            if is_refiner:
+                pipeline_kwargs.update({"force_zeros_for_empty_prompt": False})
+
+            pipe = pipeline_class(**pipeline_kwargs)
     else:
         text_config = create_ldm_bert_config(original_config)
         text_model = convert_ldm_bert_checkpoint(checkpoint, text_config)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -23,6 +23,7 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -650,6 +651,67 @@ class StableDiffusionPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -25,6 +25,7 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -718,6 +719,67 @@ class StableDiffusionImg2ImgPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -25,6 +25,7 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import FusedAttnProcessor2_0
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
@@ -844,6 +845,67 @@ class StableDiffusionInpaintPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -35,6 +35,7 @@ from ...loaders import (
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -864,6 +865,67 @@ class StableDiffusionXLImg2ImgPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py

@@ -36,6 +36,7 @@ from ...loaders import (
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -1084,6 +1085,67 @@ class StableDiffusionXLInpaintPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

src/diffusers/pipelines/wuerstchen/modeling_paella_vq_model.py

@@ -19,8 +19,8 @@ import torch
 import torch.nn as nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.autoencoders.vae import DecoderOutput, VectorQuantizer
 from ...models.modeling_utils import ModelMixin
-from ...models.vae import DecoderOutput, VectorQuantizer
 from ...models.vq_model import VQEncoderOutput
 from ...utils.accelerate_utils import apply_forward_hook
 

tests/pipelines/stable_diffusion/test_stable_diffusion.py

@@ -661,6 +661,37 @@ class StableDiffusionPipelineFastTests(
             output[0, -3:, -3:, -1], output_no_freeu[0, -3:, -3:, -1]
         ), "Disabling of FreeU should lead to results similar to the default pipeline results."
 
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = sd_pipe(**inputs).images
+        original_image_slice = image[0, -3:, -3:, -1]
+
+        sd_pipe.fuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        image = sd_pipe(**inputs).images
+        image_slice_fused = image[0, -3:, -3:, -1]
+
+        sd_pipe.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        image = sd_pipe(**inputs).images
+        image_slice_disabled = image[0, -3:, -3:, -1]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-2, rtol=1e-2
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."
+
 
 @slow
 @require_torch_gpu