Deprecate `upcast_vae` in SDXL based pipelines (#12619)

* update

* update

* Revert "update"

This reverts commit 73906381ab76da96eb8f9b841177cd4f49861eb1.

* Revert "update"

This reverts commit 21a03f93ef0fbfa5f7a7d97708f75149b1d1b3b0.

* update

* update

* update

* update

* update

examples/community/lpw_stable_diffusion_xl.py

@@ -29,7 +29,6 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
-from diffusers.models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -1328,18 +1327,8 @@ class SDXLLongPromptWeightingPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (AttnProcessor2_0, XFormersAttnProcessor),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # 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):

examples/community/mixture_tiling_sdxl.py

@@ -30,17 +30,13 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    FusedAttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler
 from diffusers.utils import (
     USE_PEFT_BACKEND,
+    deprecate,
     is_invisible_watermark_available,
     is_torch_xla_available,
     logging,
@@ -710,22 +706,8 @@ class StableDiffusionXLTilingPipeline(
         return torch.tile(weights_torch, (nbatches, self.unet.config.in_channels, 1, 1))
 
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                FusedAttnProcessor2_0,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(

examples/community/mod_controlnet_tile_sr_sdxl.py

@@ -39,16 +39,13 @@ from diffusers.models import (
     MultiControlNetModel,
     UNet2DConditionModel,
 )
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler
 from diffusers.utils import (
     USE_PEFT_BACKEND,
+    deprecate,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -1220,23 +1217,9 @@ class StableDiffusionXLControlNetTileSRPipeline(
 
         return tile_weights, tile_row_overlaps, tile_col_overlaps
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     @property
     def guidance_scale(self):

examples/community/pipeline_controlnet_xl_kolors.py

@@ -40,10 +40,6 @@ from diffusers.models import (
     MultiControlNetModel,
     UNet2DConditionModel,
 )
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -760,21 +756,8 @@ class KolorsControlNetPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     @property
     def guidance_scale(self):

examples/community/pipeline_controlnet_xl_kolors_img2img.py

@@ -40,10 +40,6 @@ from diffusers.models import (
     MultiControlNetModel,
     UNet2DConditionModel,
 )
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -930,21 +926,8 @@ class KolorsControlNetImg2ImgPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     @property
     def guidance_scale(self):

examples/community/pipeline_controlnet_xl_kolors_inpaint.py

@@ -39,10 +39,6 @@ from diffusers.models import (
     MultiControlNetModel,
     UNet2DConditionModel,
 )
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -1006,21 +1002,8 @@ class KolorsControlNetInpaintPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     @property
     def denoising_end(self):

examples/community/pipeline_demofusion_sdxl.py

@@ -16,11 +16,11 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
+    deprecate,
     is_accelerate_available,
     is_accelerate_version,
     is_invisible_watermark_available,
@@ -612,20 +612,9 @@ class DemoFusionSDXLPipeline(
 
         return image
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (AttnProcessor2_0, XFormersAttnProcessor),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)

examples/community/pipeline_faithdiff_stable_diffusion_xl.py

@@ -40,13 +40,6 @@ from diffusers.loaders import (
     UNet2DConditionLoadersMixin,
 )
 from diffusers.models import AutoencoderKL
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    FusedAttnProcessor2_0,
-    LoRAAttnProcessor2_0,
-    LoRAXFormersAttnProcessor,
-    XFormersAttnProcessor,
-)
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2D, get_down_block
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
@@ -1642,24 +1635,8 @@ class FaithDiffStableDiffusionXLPipeline(
         return latents
 
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                LoRAXFormersAttnProcessor,
-                LoRAAttnProcessor2_0,
-                FusedAttnProcessor2_0,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(

examples/community/pipeline_kolors_differential_img2img.py

@@ -22,13 +22,12 @@ from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
 from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import IPAdapterMixin, StableDiffusionXLLoraLoaderMixin
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
-from diffusers.models.attention_processor import AttnProcessor2_0, FusedAttnProcessor2_0, XFormersAttnProcessor
 from diffusers.pipelines.kolors.pipeline_output import KolorsPipelineOutput
 from diffusers.pipelines.kolors.text_encoder import ChatGLMModel
 from diffusers.pipelines.kolors.tokenizer import ChatGLMTokenizer
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.schedulers import KarrasDiffusionSchedulers
-from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils import deprecate, is_torch_xla_available, logging, replace_example_docstring
 from diffusers.utils.torch_utils import randn_tensor
 
 
@@ -709,24 +708,9 @@ class KolorsDifferentialImg2ImgPipeline(
         add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
         return add_time_ids
 
-    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                FusedAttnProcessor2_0,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(

examples/community/pipeline_kolors_inpainting.py

@@ -32,12 +32,6 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    LoRAAttnProcessor2_0,
-    LoRAXFormersAttnProcessor,
-    XFormersAttnProcessor,
-)
 from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -1008,23 +1002,8 @@ class KolorsInpaintPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                LoRAXFormersAttnProcessor,
-                LoRAAttnProcessor2_0,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(

examples/community/pipeline_sdxl_style_aligned.py

@@ -45,8 +45,6 @@ from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionMode
 from diffusers.models.attention_processor import (
     Attention,
     AttnProcessor2_0,
-    FusedAttnProcessor2_0,
-    XFormersAttnProcessor,
 )
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
@@ -1151,22 +1149,8 @@ class StyleAlignedSDXLPipeline(
         return add_time_ids
 
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                FusedAttnProcessor2_0,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     def _enable_shared_attention_processors(
         self,

examples/community/pipeline_stable_diffusion_upscale_ldm3d.py

@@ -503,24 +503,9 @@ class StableDiffusionUpscaleLDM3DPipeline(
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
-    # def upcast_vae(self):
-    #     dtype = self.vae.dtype
-    #     self.vae.to(dtype=torch.float32)
-    #     use_torch_2_0_or_xformers = isinstance(
-    #         self.vae.decoder.mid_block.attentions[0].processor,
-    #         (
-    #             AttnProcessor2_0,
-    #             XFormersAttnProcessor,
-    #             LoRAXFormersAttnProcessor,
-    #             LoRAAttnProcessor2_0,
-    #         ),
-    #     )
-    #     # if xformers or torch_2_0 is used attention block does not need
-    #     # to be in float32 which can save lots of memory
-    #     if use_torch_2_0_or_xformers:
-    #         self.vae.post_quant_conv.to(dtype)
-    #         self.vae.decoder.conv_in.to(dtype)
-    #         self.vae.decoder.mid_block.to(dtype)
+    def upcast_vae(self):
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
+        self.vae.to(dtype=torch.float32)
 
     @torch.no_grad()
     def __call__(

examples/community/pipeline_stable_diffusion_xl_attentive_eraser.py

@@ -35,12 +35,6 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    LoRAAttnProcessor2_0,
-    LoRAXFormersAttnProcessor,
-    XFormersAttnProcessor,
-)
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -1282,23 +1276,8 @@ class StableDiffusionXL_AE_Pipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                LoRAXFormersAttnProcessor,
-                LoRAAttnProcessor2_0,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # 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):

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py

@@ -25,7 +25,6 @@ from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokeniz
 from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, ControlNetModel, MultiAdapter, T2IAdapter, UNet2DConditionModel
-from diffusers.models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
@@ -34,6 +33,7 @@ from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     PIL_INTERPOLATION,
     USE_PEFT_BACKEND,
+    deprecate,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -793,20 +793,9 @@ class StableDiffusionXLControlNetAdapterPipeline(
         add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
         return add_time_ids
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (AttnProcessor2_0, XFormersAttnProcessor),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width
     def _default_height_width(self, height, width, image):

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -43,7 +43,6 @@ from diffusers.models import (
     T2IAdapter,
     UNet2DConditionModel,
 )
-from diffusers.models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
 from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
@@ -52,6 +51,7 @@ from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     PIL_INTERPOLATION,
     USE_PEFT_BACKEND,
+    deprecate,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -1130,20 +1130,9 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
 
         return add_time_ids, add_neg_time_ids
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (AttnProcessor2_0, XFormersAttnProcessor),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width
     def _default_height_width(self, height, width, image):

examples/community/pipeline_stable_diffusion_xl_differential_img2img.py

@@ -35,10 +35,6 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
@@ -848,21 +844,8 @@ class StableDiffusionXLDifferentialImg2ImgPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(

examples/community/pipeline_stable_diffusion_xl_ipex.py

@@ -32,10 +32,6 @@ from diffusers.loaders import (
     TextualInversionLoaderMixin,
 )
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers
@@ -658,23 +654,9 @@ class StableDiffusionXLPipelineIpex(
         add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
         return add_time_ids
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
+        deprecate("upcast_vae", "1.0.0", "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`")
         self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
-        )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
 
     # 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/modular_pipelines/stable_diffusion_xl/decoders.py

@@ -21,8 +21,7 @@ import torch
 from ...configuration_utils import FrozenDict
 from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKL
-from ...models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
-from ...utils import logging
+from ...utils import deprecate, logging
 from ..modular_pipeline import (
     ModularPipelineBlocks,
     PipelineState,
@@ -77,21 +76,12 @@ class StableDiffusionXLDecodeStep(ModularPipelineBlocks):
     @staticmethod
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae with self->components
     def upcast_vae(components):
-        dtype = components.vae.dtype
-        components.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            components.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
+        deprecate(
+            "upcast_vae",
+            "1.0.0",
+            "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.",
         )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            components.vae.post_quant_conv.to(dtype)
-            components.vae.decoder.conv_in.to(dtype)
-            components.vae.decoder.mid_block.to(dtype)
+        components.vae.to(dtype=torch.float32)
 
     @torch.no_grad()
     def __call__(self, components, state: PipelineState) -> PipelineState:

src/diffusers/pipelines/aura_flow/pipeline_aura_flow.py

@@ -21,10 +21,10 @@ from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...image_processor import VaeImageProcessor
 from ...loaders import AuraFlowLoraLoaderMixin
 from ...models import AuraFlowTransformer2DModel, AutoencoderKL
-from ...models.attention_processor import AttnProcessor2_0, FusedAttnProcessor2_0, XFormersAttnProcessor
 from ...schedulers import FlowMatchEulerDiscreteScheduler
 from ...utils import (
     USE_PEFT_BACKEND,
+    deprecate,
     is_torch_xla_available,
     logging,
     replace_example_docstring,
@@ -406,22 +406,12 @@ class AuraFlowPipeline(DiffusionPipeline, AuraFlowLoraLoaderMixin):
 
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
-        self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-                FusedAttnProcessor2_0,
-            ),
+        deprecate(
+            "upcast_vae",
+            "1.0.0",
+            "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.",
         )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
+        self.vae.to(dtype=torch.float32)
 
     @property
     def guidance_scale(self):

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py

@@ -36,10 +36,6 @@ from ...loaders import (
     TextualInversionLoaderMixin,
 )
 from ...models import AutoencoderKL, ControlNetModel, ImageProjection, MultiControlNetModel, UNet2DConditionModel
-from ...models.attention_processor import (
-    AttnProcessor2_0,
-    XFormersAttnProcessor,
-)
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -1132,21 +1128,12 @@ class StableDiffusionXLControlNetInpaintPipeline(
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
-        dtype = self.vae.dtype
-        self.vae.to(dtype=torch.float32)
-        use_torch_2_0_or_xformers = isinstance(
-            self.vae.decoder.mid_block.attentions[0].processor,
-            (
-                AttnProcessor2_0,
-                XFormersAttnProcessor,
-            ),
+        deprecate(
+            "upcast_vae",
+            "1.0.0",
+            "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.",
         )
-        # if xformers or torch_2_0 is used attention block does not need
-        # to be in float32 which can save lots of memory
-        if use_torch_2_0_or_xformers:
-            self.vae.post_quant_conv.to(dtype)
-            self.vae.decoder.conv_in.to(dtype)
-            self.vae.decoder.mid_block.to(dtype)
+        self.vae.to(dtype=torch.float32)
 
     @property
     def guidance_scale(self):