Merge branch 'main' into bnb-multi-backend

src/diffusers/loaders/textual_inversion.py

@@ -40,7 +40,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
     force_download = kwargs.pop("force_download", False)
     proxies = kwargs.pop("proxies", None)
     local_files_only = kwargs.pop("local_files_only", None)
-    token = kwargs.pop("token", None)
+    hf_token = kwargs.pop("hf_token", None)
     revision = kwargs.pop("revision", None)
     subfolder = kwargs.pop("subfolder", None)
     weight_name = kwargs.pop("weight_name", None)
@@ -73,7 +73,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                         force_download=force_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        token=token,
+                        token=hf_token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -93,7 +93,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     force_download=force_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    token=token,
+                    token=hf_token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -312,7 +312,7 @@ def load_textual_inversion(
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            hf_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):

src/diffusers/pipelines/pag/pipeline_pag_sana.py

@@ -16,6 +16,7 @@
 import inspect
 import re
 import urllib.parse as ul
+import warnings
 from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
@@ -41,6 +42,7 @@
     ASPECT_RATIO_1024_BIN,
 )
 from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN
+from ..sana.pipeline_sana import ASPECT_RATIO_4096_BIN
 from .pag_utils import PAGMixin
 
 
@@ -639,7 +641,7 @@ def __call__(
         negative_prompt_attention_mask: Optional[torch.Tensor] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
-        clean_caption: bool = True,
+        clean_caption: bool = False,
         use_resolution_binning: bool = True,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
@@ -755,7 +757,9 @@ def __call__(
             callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
 
         if use_resolution_binning:
-            if self.transformer.config.sample_size == 64:
+            if self.transformer.config.sample_size == 128:
+                aspect_ratio_bin = ASPECT_RATIO_4096_BIN
+            elif self.transformer.config.sample_size == 64:
                 aspect_ratio_bin = ASPECT_RATIO_2048_BIN
             elif self.transformer.config.sample_size == 32:
                 aspect_ratio_bin = ASPECT_RATIO_1024_BIN
@@ -912,7 +916,14 @@ def __call__(
             image = latents
         else:
             latents = latents.to(self.vae.dtype)
-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            try:
+                image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            except torch.cuda.OutOfMemoryError as e:
+                warnings.warn(
+                    f"{e}. \n"
+                    f"Try to use VAE tiling for large images. For example: \n"
+                    f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)"
+                )
             if use_resolution_binning:
                 image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
 

src/diffusers/pipelines/sana/pipeline_sana.py

@@ -16,6 +16,7 @@
 import inspect
 import re
 import urllib.parse as ul
+import warnings
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import torch
@@ -953,7 +954,14 @@ def __call__(
             image = latents
         else:
             latents = latents.to(self.vae.dtype)
-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            try:
+                image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            except torch.cuda.OutOfMemoryError as e:
+                warnings.warn(
+                    f"{e}. \n"
+                    f"Try to use VAE tiling for large images. For example: \n"
+                    f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)"
+                )
             if use_resolution_binning:
                 image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)