diffusers 0.32.2__py3-none-any.whl → 0.33.0__py3-none-any.whl

diffusers/pipelines/auto_pipeline.py

@@ -1,5 +1,5 @@
 # coding=utf-8
-# Copyright 2024 The HuggingFace Inc. team.
+# Copyright 2025 The HuggingFace Inc. team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -22,6 +22,7 @@ from ..models.controlnets import ControlNetUnionModel
 from ..utils import is_sentencepiece_available
 from .aura_flow import AuraFlowPipeline
 from .cogview3 import CogView3PlusPipeline
+from .cogview4 import CogView4ControlPipeline, CogView4Pipeline
 from .controlnet import (
     StableDiffusionControlNetImg2ImgPipeline,
     StableDiffusionControlNetInpaintPipeline,
@@ -33,6 +34,10 @@ from .controlnet import (
     StableDiffusionXLControlNetUnionInpaintPipeline,
     StableDiffusionXLControlNetUnionPipeline,
 )
+from .controlnet_sd3 import (
+    StableDiffusion3ControlNetInpaintingPipeline,
+    StableDiffusion3ControlNetPipeline,
+)
 from .deepfloyd_if import IFImg2ImgPipeline, IFInpaintingPipeline, IFPipeline
 from .flux import (
     FluxControlImg2ImgPipeline,
@@ -64,10 +69,12 @@ from .kandinsky2_2 import (
 )
 from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
 from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
-from .lumina import LuminaText2ImgPipeline
+from .lumina import LuminaPipeline
+from .lumina2 import Lumina2Pipeline
 from .pag import (
     HunyuanDiTPAGPipeline,
     PixArtSigmaPAGPipeline,
+    SanaPAGPipeline,
     StableDiffusion3PAGImg2ImgPipeline,
     StableDiffusion3PAGPipeline,
     StableDiffusionControlNetPAGInpaintPipeline,
@@ -82,6 +89,7 @@ from .pag import (
     StableDiffusionXLPAGPipeline,
 )
 from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
+from .sana import SanaPipeline
 from .stable_cascade import StableCascadeCombinedPipeline, StableCascadeDecoderPipeline
 from .stable_diffusion import (
     StableDiffusionImg2ImgPipeline,
@@ -116,11 +124,14 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("stable-diffusion-controlnet", StableDiffusionControlNetPipeline),
         ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetPipeline),
         ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionPipeline),
+        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetPipeline),
         ("wuerstchen", WuerstchenCombinedPipeline),
         ("cascade", StableCascadeCombinedPipeline),
         ("lcm", LatentConsistencyModelPipeline),
         ("pixart-alpha", PixArtAlphaPipeline),
         ("pixart-sigma", PixArtSigmaPipeline),
+        ("sana", SanaPipeline),
+        ("sana-pag", SanaPAGPipeline),
         ("stable-diffusion-pag", StableDiffusionPAGPipeline),
         ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGPipeline),
         ("stable-diffusion-xl-pag", StableDiffusionXLPAGPipeline),
@@ -130,8 +141,11 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("flux", FluxPipeline),
         ("flux-control", FluxControlPipeline),
         ("flux-controlnet", FluxControlNetPipeline),
-        ("lumina", LuminaText2ImgPipeline),
+        ("lumina", LuminaPipeline),
+        ("lumina2", Lumina2Pipeline),
         ("cogview3", CogView3PlusPipeline),
+        ("cogview4", CogView4Pipeline),
+        ("cogview4-control", CogView4ControlPipeline),
     ]
 )
 
@@ -170,6 +184,7 @@ AUTO_INPAINT_PIPELINES_MAPPING = OrderedDict(
         ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGInpaintPipeline),
         ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetInpaintPipeline),
         ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionInpaintPipeline),
+        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetInpaintingPipeline),
         ("stable-diffusion-xl-pag", StableDiffusionXLPAGInpaintPipeline),
         ("flux", FluxInpaintPipeline),
         ("flux-controlnet", FluxControlNetInpaintPipeline),
@@ -293,7 +308,7 @@ class AutoPipelineForText2Image(ConfigMixin):
         If you get the error message below, you need to finetune the weights for your downstream task:
 
         ```
-        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
         - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
         You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
         ```
@@ -385,7 +400,7 @@ class AutoPipelineForText2Image(ConfigMixin):
         ```py
         >>> from diffusers import AutoPipelineForText2Image
 
-        >>> pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5")
+        >>> pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
         >>> image = pipeline(prompt).images[0]
         ```
         """
@@ -448,7 +463,7 @@ class AutoPipelineForText2Image(ConfigMixin):
         >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
 
         >>> pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
-        ...     "runwayml/stable-diffusion-v1-5", requires_safety_checker=False
+        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
         ... )
 
         >>> pipe_t2i = AutoPipelineForText2Image.from_pipe(pipe_i2i)
@@ -528,7 +543,9 @@ class AutoPipelineForText2Image(ConfigMixin):
             if k not in text_2_image_kwargs
         }
 
-        missing_modules = set(expected_modules) - set(pipeline._optional_components) - set(text_2_image_kwargs.keys())
+        missing_modules = (
+            set(expected_modules) - set(text_2_image_cls._optional_components) - set(text_2_image_kwargs.keys())
+        )
 
         if len(missing_modules) > 0:
             raise ValueError(
@@ -587,7 +604,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
         If you get the error message below, you need to finetune the weights for your downstream task:
 
         ```
-        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
         - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
         You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
         ```
@@ -679,7 +696,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
         ```py
         >>> from diffusers import AutoPipelineForImage2Image
 
-        >>> pipeline = AutoPipelineForImage2Image.from_pretrained("runwayml/stable-diffusion-v1-5")
+        >>> pipeline = AutoPipelineForImage2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
         >>> image = pipeline(prompt, image).images[0]
         ```
         """
@@ -754,7 +771,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
         >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
 
         >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
-        ...     "runwayml/stable-diffusion-v1-5", requires_safety_checker=False
+        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
         ... )
 
         >>> pipe_i2i = AutoPipelineForImage2Image.from_pipe(pipe_t2i)
@@ -838,7 +855,9 @@ class AutoPipelineForImage2Image(ConfigMixin):
             if k not in image_2_image_kwargs
         }
 
-        missing_modules = set(expected_modules) - set(pipeline._optional_components) - set(image_2_image_kwargs.keys())
+        missing_modules = (
+            set(expected_modules) - set(image_2_image_cls._optional_components) - set(image_2_image_kwargs.keys())
+        )
 
         if len(missing_modules) > 0:
             raise ValueError(
@@ -896,7 +915,7 @@ class AutoPipelineForInpainting(ConfigMixin):
         If you get the error message below, you need to finetune the weights for your downstream task:
 
         ```
-        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
         - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
         You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
         ```
@@ -988,7 +1007,7 @@ class AutoPipelineForInpainting(ConfigMixin):
         ```py
         >>> from diffusers import AutoPipelineForInpainting
 
-        >>> pipeline = AutoPipelineForInpainting.from_pretrained("runwayml/stable-diffusion-v1-5")
+        >>> pipeline = AutoPipelineForInpainting.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
         >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
         ```
         """
@@ -1141,7 +1160,9 @@ class AutoPipelineForInpainting(ConfigMixin):
             if k not in inpainting_kwargs
         }
 
-        missing_modules = set(expected_modules) - set(pipeline._optional_components) - set(inpainting_kwargs.keys())
+        missing_modules = (
+            set(expected_modules) - set(inpainting_cls._optional_components) - set(inpainting_kwargs.keys())
+        )
 
         if len(missing_modules) > 0:
             raise ValueError(

diffusers/pipelines/blip_diffusion/blip_image_processing.py

@@ -1,5 +1,5 @@
 # coding=utf-8
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.

diffusers/pipelines/blip_diffusion/modeling_blip2.py

@@ -174,19 +174,16 @@ class Blip2QFormerEncoder(nn.Module):
                     )
                     use_cache = False
 
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, past_key_value, output_attentions, query_length)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer_module,
                     hidden_states,
                     attention_mask,
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
                 )
             else:
                 layer_outputs = layer_module(

diffusers/pipelines/blip_diffusion/pipeline_blip_diffusion.py

@@ -20,6 +20,7 @@ from transformers import CLIPTokenizer
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import PNDMScheduler
 from ...utils import (
+    is_torch_xla_available,
     logging,
     replace_example_docstring,
 )
@@ -30,8 +31,16 @@ from .modeling_blip2 import Blip2QFormerModel
 from .modeling_ctx_clip import ContextCLIPTextModel
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+
 EXAMPLE_DOC_STRING = """
     Examples:
         ```py
@@ -336,6 +345,9 @@ class BlipDiffusionPipeline(DiffusionPipeline):
                 latents,
             )["prev_sample"]
 
+            if XLA_AVAILABLE:
+                xm.mark_step()
+
         image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
         image = self.image_processor.postprocess(image, output_type=output_type)
 

diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -26,12 +26,19 @@ from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
 from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
-from ...utils import logging, replace_example_docstring
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from .pipeline_output import CogVideoXPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -183,14 +190,12 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
             tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
         )
         self.vae_scale_factor_spatial = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+            2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         )
         self.vae_scale_factor_temporal = (
-            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
-        )
-        self.vae_scaling_factor_image = (
-            self.vae.config.scaling_factor if hasattr(self, "vae") and self.vae is not None else 0.7
+            self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
         )
+        self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
 
         self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
 
@@ -489,6 +494,10 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
     def attention_kwargs(self):
         return self._attention_kwargs
 
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
     @property
     def interrupt(self):
         return self._interrupt
@@ -622,6 +631,7 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         )
         self._guidance_scale = guidance_scale
         self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
         self._interrupt = False
 
         # 2. Default call parameters
@@ -700,6 +710,7 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 if self.interrupt:
                     continue
 
+                self._current_timestep = t
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
@@ -755,6 +766,11 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
         if not output_type == "latent":
             # Discard any padding frames that were added for CogVideoX 1.5
             latents = latents[:, additional_frames:]

diffusers/pipelines/cogvideo/pipeline_cogvideox_fun_control.py

@@ -27,12 +27,19 @@ from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
 from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import logging, replace_example_docstring
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from .pipeline_output import CogVideoXPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -190,14 +197,12 @@ class CogVideoXFunControlPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
             tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
         )
         self.vae_scale_factor_spatial = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+            2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         )
         self.vae_scale_factor_temporal = (
-            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
-        )
-        self.vae_scaling_factor_image = (
-            self.vae.config.scaling_factor if hasattr(self, "vae") and self.vae is not None else 0.7
+            self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
         )
+        self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
 
         self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
 
@@ -535,6 +540,10 @@ class CogVideoXFunControlPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
     def attention_kwargs(self):
         return self._attention_kwargs
 
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
     @property
     def interrupt(self):
         return self._interrupt
@@ -675,6 +684,7 @@ class CogVideoXFunControlPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         )
         self._guidance_scale = guidance_scale
         self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
         self._interrupt = False
 
         # 2. Default call parameters
@@ -761,6 +771,7 @@ class CogVideoXFunControlPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 if self.interrupt:
                     continue
 
+                self._current_timestep = t
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
@@ -810,6 +821,11 @@ class CogVideoXFunControlPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
         if not output_type == "latent":
             video = self.decode_latents(latents)
             video = self.video_processor.postprocess_video(video=video, output_type=output_type)

diffusers/pipelines/cogvideo/pipeline_cogvideox_image2video.py

@@ -29,6 +29,7 @@ from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
 from ...utils import (
+    is_torch_xla_available,
     logging,
     replace_example_docstring,
 )
@@ -37,6 +38,13 @@ from ...video_processor import VideoProcessor
 from .pipeline_output import CogVideoXPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -203,14 +211,12 @@ class CogVideoXImageToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
             scheduler=scheduler,
         )
         self.vae_scale_factor_spatial = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+            2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         )
         self.vae_scale_factor_temporal = (
-            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
-        )
-        self.vae_scaling_factor_image = (
-            self.vae.config.scaling_factor if hasattr(self, "vae") and self.vae is not None else 0.7
+            self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
         )
+        self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
 
         self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
 
@@ -585,6 +591,10 @@ class CogVideoXImageToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
     def attention_kwargs(self):
         return self._attention_kwargs
 
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
     @property
     def interrupt(self):
         return self._interrupt
@@ -722,6 +732,7 @@ class CogVideoXImageToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
             negative_prompt_embeds=negative_prompt_embeds,
         )
         self._guidance_scale = guidance_scale
+        self._current_timestep = None
         self._attention_kwargs = attention_kwargs
         self._interrupt = False
 
@@ -809,6 +820,7 @@ class CogVideoXImageToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
                 if self.interrupt:
                     continue
 
+                self._current_timestep = t
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
@@ -868,6 +880,11 @@ class CogVideoXImageToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
         if not output_type == "latent":
             # Discard any padding frames that were added for CogVideoX 1.5
             latents = latents[:, additional_frames:]

diffusers/pipelines/cogvideo/pipeline_cogvideox_video2video.py

@@ -27,12 +27,19 @@ from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
 from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
-from ...utils import logging, replace_example_docstring
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from .pipeline_output import CogVideoXPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -206,14 +213,12 @@ class CogVideoXVideoToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
         )
 
         self.vae_scale_factor_spatial = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+            2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         )
         self.vae_scale_factor_temporal = (
-            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
-        )
-        self.vae_scaling_factor_image = (
-            self.vae.config.scaling_factor if hasattr(self, "vae") and self.vae is not None else 0.7
+            self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
         )
+        self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
 
         self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
 
@@ -559,6 +564,10 @@ class CogVideoXVideoToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
     def attention_kwargs(self):
         return self._attention_kwargs
 
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
     @property
     def interrupt(self):
         return self._interrupt
@@ -695,6 +704,7 @@ class CogVideoXVideoToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
         )
         self._guidance_scale = guidance_scale
         self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
         self._interrupt = False
 
         # 2. Default call parameters
@@ -781,6 +791,7 @@ class CogVideoXVideoToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
                 if self.interrupt:
                     continue
 
+                self._current_timestep = t
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
@@ -836,6 +847,11 @@ class CogVideoXVideoToVideoPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin)
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
         if not output_type == "latent":
             video = self.decode_latents(latents)
             video = self.video_processor.postprocess_video(video=video, output_type=output_type)

diffusers/pipelines/cogview3/pipeline_cogview3plus.py

@@ -24,11 +24,18 @@ from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKL, CogView3PlusTransformer2DModel
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
-from ...utils import logging, replace_example_docstring
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
 from .pipeline_output import CogView3PipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -153,9 +160,7 @@ class CogView3PlusPipeline(DiffusionPipeline):
         self.register_modules(
             tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
         )
-        self.vae_scale_factor = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
-        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
 
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
 
@@ -656,6 +661,9 @@ class CogView3PlusPipeline(DiffusionPipeline):
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
                 0

diffusers/pipelines/cogview4/__init__.py

@@ -0,0 +1,49 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_additional_imports = {}
+_import_structure = {"pipeline_output": ["CogView4PlusPipelineOutput"]}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_cogview4"] = ["CogView4Pipeline"]
+    _import_structure["pipeline_cogview4_control"] = ["CogView4ControlPipeline"]
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .pipeline_cogview4 import CogView4Pipeline
+        from .pipeline_cogview4_control import CogView4ControlPipeline
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)
+    for name, value in _additional_imports.items():
+        setattr(sys.modules[__name__], name, value)