Merge branch 'main' into patch-1

Author: yiyixuxu

docs/source/en/api/pipelines/pag.md

@@ -53,6 +53,11 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
 	- all
 	- __call__
 
+## StableDiffusionPAGImg2ImgPipeline
+[[autodoc]] StableDiffusionPAGImg2ImgPipeline
+	- all
+	- __call__
+
 ## StableDiffusionControlNetPAGPipeline
 [[autodoc]] StableDiffusionControlNetPAGPipeline
 

docs/source/en/tutorials/using_peft_for_inference.md

@@ -75,6 +75,12 @@ image
 
 ![pixel-art](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_12_1.png)
 
+<Tip>
+
+By default, if the most up-to-date versions of PEFT and Transformers are detected, `low_cpu_mem_usage` is set to `True` to speed up the loading time of LoRA checkpoints. 
+
+</Tip>
+
 ## Merge adapters
 
 You can also merge different adapter checkpoints for inference to blend their styles together.

src/diffusers/__init__.py

@@ -344,6 +344,7 @@
             "StableDiffusionLatentUpscalePipeline",
             "StableDiffusionLDM3DPipeline",
             "StableDiffusionModelEditingPipeline",
+            "StableDiffusionPAGImg2ImgPipeline",
             "StableDiffusionPAGPipeline",
             "StableDiffusionPanoramaPipeline",
             "StableDiffusionParadigmsPipeline",
@@ -795,6 +796,7 @@
             StableDiffusionLatentUpscalePipeline,
             StableDiffusionLDM3DPipeline,
             StableDiffusionModelEditingPipeline,
+            StableDiffusionPAGImg2ImgPipeline,
             StableDiffusionPAGPipeline,
             StableDiffusionPanoramaPipeline,
             StableDiffusionParadigmsPipeline,

src/diffusers/loaders/lora_pipeline.py

@@ -115,6 +115,9 @@ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict
                 `default_{i}` where i is the total number of adapters being loaded.
             weight_name (`str`, *optional*, defaults to None):
                 Name of the serialized state dict file.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
 
         Example:
 
@@ -142,8 +145,14 @@ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict
         adapter_name = kwargs.pop("adapter_name", None)
         _pipeline = kwargs.pop("_pipeline", None)
         network_alphas = kwargs.pop("network_alphas", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", False)
         allow_pickle = False
 
+        if low_cpu_mem_usage and is_peft_version("<=", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
         if use_safetensors is None:
             use_safetensors = True
             allow_pickle = True
@@ -209,6 +218,7 @@ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict
                 network_alphas=network_alphas,
                 adapter_name=adapter_name,
                 _pipeline=_pipeline,
+                low_cpu_mem_usage=low_cpu_mem_usage,
             )
         else:
             raise ValueError(
@@ -268,7 +278,9 @@ def _process_custom_diffusion(self, state_dict):
 
         return attn_processors
 
-    def _process_lora(self, state_dict, unet_identifier_key, network_alphas, adapter_name, _pipeline):
+    def _process_lora(
+        self, state_dict, unet_identifier_key, network_alphas, adapter_name, _pipeline, low_cpu_mem_usage
+    ):
         # This method does the following things:
         # 1. Filters the `state_dict` with keys matching  `unet_identifier_key` when using the non-legacy
         #    format. For legacy format no filtering is applied.
@@ -335,9 +347,12 @@ def _process_lora(self, state_dict, unet_identifier_key, network_alphas, adapter
             # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
             # otherwise loading LoRA weights will lead to an error
             is_model_cpu_offload, is_sequential_cpu_offload = self._optionally_disable_offloading(_pipeline)
+            peft_kwargs = {}
+            if is_peft_version(">=", "0.13.1"):
+                peft_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
 
-            inject_adapter_in_model(lora_config, self, adapter_name=adapter_name)
-            incompatible_keys = set_peft_model_state_dict(self, state_dict, adapter_name)
+            inject_adapter_in_model(lora_config, self, adapter_name=adapter_name, **peft_kwargs)
+            incompatible_keys = set_peft_model_state_dict(self, state_dict, adapter_name, **peft_kwargs)
 
             if incompatible_keys is not None:
                 # check only for unexpected keys

src/diffusers/loaders/unet.py

@@ -164,6 +164,7 @@
             "HunyuanDiTPAGPipeline",
             "StableDiffusion3PAGPipeline",
             "StableDiffusionPAGPipeline",
+            "StableDiffusionPAGImg2ImgPipeline",
             "StableDiffusionControlNetPAGPipeline",
             "StableDiffusionXLPAGPipeline",
             "StableDiffusionXLPAGInpaintPipeline",
@@ -569,6 +570,7 @@
             StableDiffusion3PAGPipeline,
             StableDiffusionControlNetPAGInpaintPipeline,
             StableDiffusionControlNetPAGPipeline,
+            StableDiffusionPAGImg2ImgPipeline,
             StableDiffusionPAGPipeline,
             StableDiffusionXLControlNetPAGImg2ImgPipeline,
             StableDiffusionXLControlNetPAGPipeline,

src/diffusers/pipelines/__init__.py

@@ -63,6 +63,7 @@
     StableDiffusion3PAGPipeline,
     StableDiffusionControlNetPAGInpaintPipeline,
     StableDiffusionControlNetPAGPipeline,
+    StableDiffusionPAGImg2ImgPipeline,
     StableDiffusionPAGPipeline,
     StableDiffusionXLControlNetPAGImg2ImgPipeline,
     StableDiffusionXLControlNetPAGPipeline,
@@ -131,6 +132,7 @@
         ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline),
         ("kandinsky3", Kandinsky3Img2ImgPipeline),
         ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
+        ("stable-diffusion-pag", StableDiffusionPAGImg2ImgPipeline),
         ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
         ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
         ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline),

src/diffusers/pipelines/auto_pipeline.py

@@ -893,6 +893,10 @@ def cross_attention_kwargs(self):
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -1089,6 +1093,7 @@ def __call__(
         self._guidance_scale = guidance_scale
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
 
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
@@ -1235,6 +1240,9 @@ def __call__(
         is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
                 # Relevant thread:
                 # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
                 if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:

src/diffusers/pipelines/controlnet/pipeline_controlnet.py

@@ -891,6 +891,10 @@ def cross_attention_kwargs(self):
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -1081,6 +1085,7 @@ def __call__(
         self._guidance_scale = guidance_scale
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
 
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
@@ -1211,6 +1216,9 @@ def __call__(
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

src/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py

@@ -976,6 +976,10 @@ def cross_attention_kwargs(self):
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -1191,6 +1195,7 @@ def __call__(
         self._guidance_scale = guidance_scale
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
 
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
@@ -1375,6 +1380,9 @@ def __call__(
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)