obscura/diffusers
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

update

Browse Source
This commit is contained in:
Dhruv Nair
2025-11-10 08:07:26 +01:00
parent a9cb08af39
commit 21a03f93ef
48 changed files with 1335 additions and 2065 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/community/lpw_stable_diffusion_xl.py
+2 -14
View File
@@ -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 +1327,7 @@ class SDXLLongPromptWeightingPipeline(
# 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),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# 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):
@@ -1921,7 +1909,7 @@ class SDXLLongPromptWeightingPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
examples/community/masked_stable_diffusion_xl_img2img.py
+1 -1
View File
@@ -481,7 +481,7 @@ class MaskedStableDiffusionXLImg2ImgPipeline(StableDiffusionXLImg2ImgPipeline):
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
examples/community/mixture_tiling_sdxl.py
+1218 -1237
View File
File diff suppressed because it is too large Load Diff
examples/community/mod_controlnet_tile_sr_sdxl.py
+3 -20
View File
@@ -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,
@@ -1222,21 +1219,7 @@ class StableDiffusionXLControlNetTileSRPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1820,7 +1803,7 @@ class StableDiffusionXLControlNetTileSRPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
examples/community/pipeline_controlnet_xl_kolors.py
+2 -20
View File
@@ -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 +756,7 @@ class KolorsControlNetPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1331,7 +1313,7 @@ class KolorsControlNetPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
latents = latents / self.vae.config.scaling_factor
examples/community/pipeline_controlnet_xl_kolors_img2img.py
+2 -20
View File
@@ -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 +926,7 @@ class KolorsControlNetImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1533,7 +1515,7 @@ class KolorsControlNetImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
latents = latents / self.vae.config.scaling_factor
examples/community/pipeline_controlnet_xl_kolors_inpaint.py
+2 -20
View File
@@ -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 +1002,7 @@ class KolorsControlNetInpaintPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def denoising_end(self):
@@ -1847,7 +1829,7 @@ class KolorsControlNetInpaintPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
latents = latents / self.vae.config.scaling_factor
examples/community/pipeline_demofusion_sdxl.py
+4 -15
View File
@@ -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,
@@ -614,18 +614,7 @@ class DemoFusionSDXLPipeline(
# 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),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
@@ -997,7 +986,7 @@ class DemoFusionSDXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
print("### Phase 1 Decoding ###")
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
@@ -1257,7 +1246,7 @@ class DemoFusionSDXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
print("### Phase {} Decoding ###".format(current_scale_num))
examples/community/pipeline_faithdiff_stable_diffusion_xl.py
+3 -27
View File
@@ -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
@@ -1637,24 +1630,7 @@ class FaithDiffStableDiffusionXLPipeline(
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,
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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1782,7 +1758,7 @@ class FaithDiffStableDiffusionXLPipeline(
# needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
# if needs_upcasting:
# image = image.float()
# self.upcast_vae()
# self.vae.to(torch.float32)
self.unet.denoise_encoder.to(device=image.device, dtype=image.dtype)
image_latents = self.unet.denoise_encoder(image)
self.unet.denoise_encoder.to("cpu")
@@ -2235,7 +2211,7 @@ class FaithDiffStableDiffusionXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
examples/community/pipeline_kolors_differential_img2img.py
+3 -19
View File
@@ -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
@@ -711,22 +710,7 @@ class KolorsDifferentialImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1252,7 +1236,7 @@ class KolorsDifferentialImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
examples/community/pipeline_kolors_inpainting.py
+2 -24
View File
@@ -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 +1002,7 @@ class KolorsInpaintPipeline(
# 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,
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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1680,7 +1658,7 @@ class KolorsInpaintPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
examples/community/pipeline_sdxl_style_aligned.py
+2 -19
View File
@@ -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 +1149,7 @@ class StyleAlignedSDXLPipeline(
return add_time_ids
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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
def _enable_shared_attention_processors(
self,
@@ -1892,7 +1875,7 @@ class StyleAlignedSDXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
examples/community/pipeline_stable_diffusion_upscale_ldm3d.py
+1 -1
View File
@@ -736,7 +736,7 @@ class StableDiffusionUpscaleLDM3DPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.scaling_factor, return_dict=False)[0]
examples/community/pipeline_stable_diffusion_xl_attentive_eraser.py
+2 -24
View File
@@ -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 +1276,7 @@ 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
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)
deprecate("`upcast_vae` is deprecated")
# 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):
@@ -2274,7 +2252,7 @@ class StableDiffusionXL_AE_Pipeline(
latents = latents[-1:]
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py
+3 -14
View File
@@ -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,
@@ -795,18 +795,7 @@ class StableDiffusionXLControlNetAdapterPipeline(
# 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),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width
def _default_height_width(self, height, width, image):
@@ -1385,7 +1374,7 @@ class StableDiffusionXLControlNetAdapterPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py
+3 -14
View File
@@ -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,
@@ -1132,18 +1132,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
# 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),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width
def _default_height_width(self, height, width, image):
@@ -1825,7 +1814,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
# make sure the VAE is in float32 mode, as it overflows in float16
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if output_type != "latent":
examples/community/pipeline_stable_diffusion_xl_differential_img2img.py
+2 -20
View File
@@ -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 +844,7 @@ class StableDiffusionXLDifferentialImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1421,7 +1403,7 @@ class StableDiffusionXLDifferentialImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
examples/community/pipeline_stable_diffusion_xl_instandid_img2img.py
+1 -1
View File
@@ -1050,7 +1050,7 @@ class StableDiffusionXLInstantIDImg2ImgPipeline(StableDiffusionXLControlNetImg2I
# make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
examples/community/pipeline_stable_diffusion_xl_instantid.py
+1 -1
View File
@@ -1039,7 +1039,7 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
# make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
examples/community/pipeline_stable_diffusion_xl_ipex.py
+3 -21
View File
@@ -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
@@ -660,21 +656,7 @@ class StableDiffusionXLPipelineIpex(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# 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):
@@ -1150,7 +1132,7 @@ class StableDiffusionXLPipelineIpex(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
@@ -1363,7 +1345,7 @@ class StableDiffusionXLPipelineIpex(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# cast back to fp16 if needed
examples/community/stable_diffusion_xl_controlnet_reference.py
+2 -2
View File
@@ -195,7 +195,7 @@ class StableDiffusionXLControlNetReferencePipeline(StableDiffusionXLControlNetPi
refimage = refimage.to(device=device)
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
refimage = refimage.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if refimage.dtype != self.vae.dtype:
refimage = refimage.to(dtype=self.vae.dtype)
@@ -1326,7 +1326,7 @@ class StableDiffusionXLControlNetReferencePipeline(StableDiffusionXLControlNetPi
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
examples/community/stable_diffusion_xl_reference.py
+2 -2
View File
@@ -141,7 +141,7 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
refimage = refimage.to(device=device)
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
refimage = refimage.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if refimage.dtype != self.vae.dtype:
refimage = refimage.to(dtype=self.vae.dtype)
@@ -1156,7 +1156,7 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/animatediff/pipeline_animatediff_sdxl.py
+3 -22
View File
@@ -32,11 +32,6 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, MotionAdapter, UNet2DConditionModel, UNetMotionModel
from ...models.attention_processor import (
AttnProcessor2_0,
FusedAttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import (
DDIMScheduler,
@@ -48,6 +43,7 @@ from ...schedulers import (
)
from ...utils import (
USE_PEFT_BACKEND,
deprecate,
is_torch_xla_available,
logging,
replace_example_docstring,
@@ -783,22 +779,7 @@ class AnimateDiffSDXLPipeline(
return add_time_ids
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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1286,7 +1267,7 @@ class AnimateDiffSDXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# 10. Post processing
src/diffusers/pipelines/aura_flow/pipeline_aura_flow.py
+3 -18
View File
@@ -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 +406,7 @@ 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -663,7 +648,7 @@ class AuraFlowPipeline(DiffusionPipeline, AuraFlowLoraLoaderMixin):
# make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
image = self.image_processor.postprocess(image, output_type=output_type)
src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py
+2 -20
View File
@@ -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 +1128,7 @@ 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1850,7 +1832,7 @@ class StableDiffusionXLControlNetInpaintPipeline(
# make sure the VAE is in float32 mode, as it overflows in float16
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# If we do sequential model offloading, let's offload unet and controlnet
src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
+2 -20
View File
@@ -39,10 +39,6 @@ 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 (
@@ -929,21 +925,7 @@ class StableDiffusionXLControlNetPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1574,7 +1556,7 @@ class StableDiffusionXLControlNetPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py
+2 -20
View File
@@ -39,10 +39,6 @@ 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 (
@@ -1044,21 +1040,7 @@ class StableDiffusionXLControlNetImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1639,7 +1621,7 @@ class StableDiffusionXLControlNetImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
src/diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py
+2 -20
View File
@@ -41,10 +41,6 @@ from ...models import (
MultiControlNetUnionModel,
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 (
@@ -1111,21 +1107,7 @@ class StableDiffusionXLControlNetUnionInpaintPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1877,7 +1859,7 @@ class StableDiffusionXLControlNetUnionInpaintPipeline(
# make sure the VAE is in float32 mode, as it overflows in float16
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# If we do sequential model offloading, let's offload unet and controlnet
src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl.py
+3 -20
View File
@@ -44,14 +44,11 @@ from ...models import (
MultiControlNetUnionModel,
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 (
USE_PEFT_BACKEND,
deprecate,
logging,
replace_example_docstring,
scale_lora_layers,
@@ -904,21 +901,7 @@ class StableDiffusionXLControlNetUnionPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1578,7 +1561,7 @@ class StableDiffusionXLControlNetUnionPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py
+2 -20
View File
@@ -44,10 +44,6 @@ from ...models import (
MultiControlNetUnionModel,
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 (
@@ -1032,21 +1028,7 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1683,7 +1665,7 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs_sd_xl.py
+4 -21
View File
@@ -32,14 +32,11 @@ from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
from ...models import AutoencoderKL, ControlNetXSAdapter, UNet2DConditionModel, UNetControlNetXSModel
from ...models.attention_processor import (
AttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import (
USE_PEFT_BACKEND,
deprecate,
logging,
replace_example_docstring,
scale_lora_layers,
@@ -685,21 +682,7 @@ class StableDiffusionXLControlNetXSPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_scale
@@ -1095,7 +1078,7 @@ class StableDiffusionXLControlNetXSPipeline(
# manually for max memory savings
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if not output_type == "latent":
@@ -1103,7 +1086,7 @@ class StableDiffusionXLControlNetXSPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
src/diffusers/pipelines/kolors/pipeline_kolors.py
+3 -19
View File
@@ -21,9 +21,8 @@ from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin
from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
from ...models.attention_processor import AttnProcessor2_0, FusedAttnProcessor2_0, XFormersAttnProcessor
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import is_torch_xla_available, logging, replace_example_docstring
from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from .pipeline_output import KolorsPipelineOutput
@@ -580,22 +579,7 @@ class KolorsPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffusionLor
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1042,7 +1026,7 @@ class KolorsPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffusionLor
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/kolors/pipeline_kolors_img2img.py
+3 -19
View File
@@ -22,9 +22,8 @@ from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import IPAdapterMixin, StableDiffusionXLLoraLoaderMixin
from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
from ...models.attention_processor import AttnProcessor2_0, FusedAttnProcessor2_0, XFormersAttnProcessor
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import is_torch_xla_available, logging, replace_example_docstring
from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring
from ...utils.torch_utils import empty_device_cache, randn_tensor
from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from .pipeline_output import KolorsPipelineOutput
@@ -708,22 +707,7 @@ class KolorsImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffu
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1222,7 +1206,7 @@ class KolorsImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffu
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py
+4 -20
View File
@@ -37,8 +37,6 @@ from ...models import AutoencoderKL, UNet2DConditionModel
from ...models.attention_processor import (
Attention,
AttnProcessor,
AttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import DDIMScheduler, DPMSolverMultistepScheduler
@@ -688,21 +686,7 @@ class LEditsPPPipelineStableDiffusionXL(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1421,7 +1405,7 @@ class LEditsPPPipelineStableDiffusionXL(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
@@ -1472,7 +1456,7 @@ class LEditsPPPipelineStableDiffusionXL(
if needs_upcasting:
image = image.float()
self.upcast_vae()
self.vae.to(torch.float32)
x0 = self.vae.encode(image).latent_dist.mode()
x0 = x0.to(dtype)
@@ -1638,7 +1622,7 @@ class LEditsPPPipelineStableDiffusionXL(
# autoencoder reconstruction
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
self.vae.to(torch.float32)
x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image_rec = self.vae.decode(
x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator
src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl.py
+3 -20
View File
@@ -39,14 +39,11 @@ 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 (
USE_PEFT_BACKEND,
deprecate,
logging,
replace_example_docstring,
scale_lora_layers,
@@ -936,21 +933,7 @@ class StableDiffusionXLControlNetPAGPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1586,7 +1569,7 @@ class StableDiffusionXLControlNetPAGPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py
+3 -20
View File
@@ -39,14 +39,11 @@ 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 (
USE_PEFT_BACKEND,
deprecate,
logging,
replace_example_docstring,
scale_lora_layers,
@@ -1049,21 +1046,7 @@ class StableDiffusionXLControlNetPAGImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -1655,7 +1638,7 @@ class StableDiffusionXLControlNetPAGImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
src/diffusers/pipelines/pag/pipeline_pag_kolors.py
+3 -19
View File
@@ -21,9 +21,8 @@ from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import IPAdapterMixin, StableDiffusionXLLoraLoaderMixin
from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
from ...models.attention_processor import AttnProcessor2_0, FusedAttnProcessor2_0, XFormersAttnProcessor
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import is_torch_xla_available, logging, replace_example_docstring
from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring
from ...utils.torch_utils import randn_tensor
from ..kolors.pipeline_output import KolorsPipelineOutput
from ..kolors.text_encoder import ChatGLMModel
@@ -598,22 +597,7 @@ class KolorsPAGPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1105,7 +1089,7 @@ class KolorsPAGPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/pag/pipeline_pag_sd_xl.py
+3 -22
View File
@@ -32,15 +32,11 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
from ...models.attention_processor import (
AttnProcessor2_0,
FusedAttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import (
USE_PEFT_BACKEND,
deprecate,
is_invisible_watermark_available,
is_torch_xla_available,
logging,
@@ -762,22 +758,7 @@ class StableDiffusionXLPAGPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1305,7 +1286,7 @@ class StableDiffusionXLPAGPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py
+3 -20
View File
@@ -34,14 +34,11 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, 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 (
USE_PEFT_BACKEND,
deprecate,
is_invisible_watermark_available,
is_torch_xla_available,
logging,
@@ -910,21 +907,7 @@ class StableDiffusionXLPAGImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1501,7 +1484,7 @@ class StableDiffusionXLPAGImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/pag/pipeline_pag_sd_xl_inpaint.py
+3 -20
View File
@@ -34,14 +34,11 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, 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 (
USE_PEFT_BACKEND,
deprecate,
is_invisible_watermark_available,
is_torch_xla_available,
logging,
@@ -1001,21 +998,7 @@ class StableDiffusionXLPAGInpaintPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1729,7 +1712,7 @@ class StableDiffusionXLPAGInpaintPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
+2 -20
View File
@@ -24,10 +24,6 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
from ...models import AutoencoderKL, UNet2DConditionModel
from ...models.attention_processor import (
AttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers
from ...utils import (
@@ -529,21 +525,7 @@ class StableDiffusionUpscalePipeline(
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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@torch.no_grad()
def __call__(
@@ -791,7 +773,7 @@ class StableDiffusionUpscalePipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
# Ensure latents are always the same type as the VAE
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
src/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py
+3 -22
View File
@@ -33,15 +33,11 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, UNet2DConditionModel
from ...models.attention_processor import (
AttnProcessor2_0,
FusedAttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler
from ...utils import (
USE_PEFT_BACKEND,
deprecate,
logging,
replace_example_docstring,
scale_lora_layers,
@@ -545,22 +541,7 @@ class StableDiffusionXLKDiffusionPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@property
def guidance_scale(self):
@@ -880,7 +861,7 @@ class StableDiffusionXLKDiffusionPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
+2 -22
View File
@@ -33,11 +33,6 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
from ...models.attention_processor import (
AttnProcessor2_0,
FusedAttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import (
@@ -750,22 +745,7 @@ class StableDiffusionXLPipeline(
return add_time_ids
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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1267,7 +1247,7 @@ class StableDiffusionXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
+2 -20
View File
@@ -34,10 +34,6 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, 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 (
@@ -897,21 +893,7 @@ class StableDiffusionXLImg2ImgPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1458,7 +1440,7 @@ class StableDiffusionXLImg2ImgPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py
+2 -20
View File
@@ -35,10 +35,6 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, 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 (
@@ -1002,21 +998,7 @@ class StableDiffusionXLInpaintPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
@@ -1692,7 +1674,7 @@ class StableDiffusionXLInpaintPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py
+3 -23
View File
@@ -22,11 +22,6 @@ from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokeniz
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
from ...models import AutoencoderKL, UNet2DConditionModel
from ...models.attention_processor import (
AttnProcessor2_0,
FusedAttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import (
@@ -531,7 +526,7 @@ class StableDiffusionXLInstructPix2PixPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
image = image.float()
self.upcast_vae()
self.vae.to(torch.float32)
image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax")
@@ -590,22 +585,7 @@ class StableDiffusionXLInstructPix2PixPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
@@ -955,7 +935,7 @@ class StableDiffusionXLInstructPix2PixPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py
+3 -20
View File
@@ -34,15 +34,12 @@ from ...loaders import (
TextualInversionLoaderMixin,
)
from ...models import AutoencoderKL, ImageProjection, MultiAdapter, T2IAdapter, 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 (
PIL_INTERPOLATION,
USE_PEFT_BACKEND,
deprecate,
is_torch_xla_available,
logging,
replace_example_docstring,
@@ -779,21 +776,7 @@ class StableDiffusionXLAdapterPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width
def _default_height_width(self, height, width, image):
@@ -1288,7 +1271,7 @@ class StableDiffusionXLAdapterPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py
+3 -22
View File
@@ -19,16 +19,12 @@ from transformers import (
from ...image_processor import VaeImageProcessor
from ...loaders import StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
from ...models import AutoencoderKL, UNet2DConditionModel
from ...models.attention_processor import (
AttnProcessor2_0,
FusedAttnProcessor2_0,
XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import (
USE_PEFT_BACKEND,
BaseOutput,
deprecate,
is_invisible_watermark_available,
logging,
scale_lora_layers,
@@ -457,22 +453,7 @@ class TextToVideoZeroSDXLPipeline(
# 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,
),
)
# 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)
deprecate("`upcast_vae` is deprecated")
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
def _get_add_time_ids(
@@ -1326,7 +1307,7 @@ class TextToVideoZeroSDXLPipeline(
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
self.vae.to(torch.float32)
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]