add support for transformers==4.25.1

add fallback for when quick model creation fails

modules/sd_disable_initialization.py

@@ -30,29 +30,52 @@ class DisableInitialization:
         def CLIPTextModel_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs):
             return self.CLIPTextModel_from_pretrained(None, *model_args, config=pretrained_model_name_or_path, state_dict={}, **kwargs)
 
-        def transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs):
+        def transformers_modeling_utils_load_pretrained_model(*args, **kwargs):
+            args = args[0:3] + ('/', ) + args[4:]  # resolved_archive_file; must set it to something to prevent what seems to be a bug
+            return self.transformers_modeling_utils_load_pretrained_model(*args, **kwargs)
+
+        def transformers_utils_hub_get_file_from_cache(original, url, *args, **kwargs):
 
             # this file is always 404, prevent making request
             if url == 'https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/added_tokens.json':
                 raise transformers.utils.hub.EntryNotFoundError
 
             try:
-                return self.transformers_utils_hub_get_from_cache(url, *args, local_files_only=True, **kwargs)
+                return original(url, *args, local_files_only=True, **kwargs)
             except Exception as e:
-                return self.transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs)
+                return original(url, *args, local_files_only=False, **kwargs)
+
+        def transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs):
+            return transformers_utils_hub_get_file_from_cache(self.transformers_utils_hub_get_from_cache, url, *args, **kwargs)
+
+        def transformers_tokenization_utils_base_cached_file(url, *args, local_files_only=False, **kwargs):
+            return transformers_utils_hub_get_file_from_cache(self.transformers_tokenization_utils_base_cached_file, url, *args, **kwargs)
+
+        def transformers_configuration_utils_cached_file(url, *args, local_files_only=False, **kwargs):
+            return transformers_utils_hub_get_file_from_cache(self.transformers_configuration_utils_cached_file, url, *args, **kwargs)
 
         self.init_kaiming_uniform = torch.nn.init.kaiming_uniform_
         self.init_no_grad_normal = torch.nn.init._no_grad_normal_
         self.init_no_grad_uniform_ = torch.nn.init._no_grad_uniform_
         self.create_model_and_transforms = open_clip.create_model_and_transforms
         self.CLIPTextModel_from_pretrained = ldm.modules.encoders.modules.CLIPTextModel.from_pretrained
-        self.transformers_utils_hub_get_from_cache = transformers.utils.hub.get_from_cache
+        self.transformers_modeling_utils_load_pretrained_model = getattr(transformers.modeling_utils.PreTrainedModel, '_load_pretrained_model', None)
+        self.transformers_tokenization_utils_base_cached_file = getattr(transformers.tokenization_utils_base, 'cached_file', None)
+        self.transformers_configuration_utils_cached_file = getattr(transformers.configuration_utils, 'cached_file', None)
+        self.transformers_utils_hub_get_from_cache = getattr(transformers.utils.hub, 'get_from_cache', None)
 
         torch.nn.init.kaiming_uniform_ = do_nothing
         torch.nn.init._no_grad_normal_ = do_nothing
         torch.nn.init._no_grad_uniform_ = do_nothing
         open_clip.create_model_and_transforms = create_model_and_transforms_without_pretrained
         ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = CLIPTextModel_from_pretrained
+        if self.transformers_modeling_utils_load_pretrained_model is not None:
+            transformers.modeling_utils.PreTrainedModel._load_pretrained_model = transformers_modeling_utils_load_pretrained_model
+        if self.transformers_tokenization_utils_base_cached_file is not None:
+            transformers.tokenization_utils_base.cached_file = transformers_tokenization_utils_base_cached_file
+        if self.transformers_configuration_utils_cached_file is not None:
+            transformers.configuration_utils.cached_file = transformers_configuration_utils_cached_file
+        if self.transformers_utils_hub_get_from_cache is not None:
             transformers.utils.hub.get_from_cache = transformers_utils_hub_get_from_cache
 
     def __exit__(self, exc_type, exc_val, exc_tb):
@@ -61,5 +84,12 @@ class DisableInitialization:
         torch.nn.init._no_grad_uniform_ = self.init_no_grad_uniform_
         open_clip.create_model_and_transforms = self.create_model_and_transforms
         ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = self.CLIPTextModel_from_pretrained
+        if self.transformers_modeling_utils_load_pretrained_model is not None:
+            transformers.modeling_utils.PreTrainedModel._load_pretrained_model = self.transformers_modeling_utils_load_pretrained_model
+        if self.transformers_tokenization_utils_base_cached_file is not None:
+            transformers.utils.hub.cached_file = self.transformers_tokenization_utils_base_cached_file
+        if self.transformers_configuration_utils_cached_file is not None:
+            transformers.utils.hub.cached_file = self.transformers_configuration_utils_cached_file
+        if self.transformers_utils_hub_get_from_cache is not None:
             transformers.utils.hub.get_from_cache = self.transformers_utils_hub_get_from_cache
 

modules/sd_models.py

@@ -14,7 +14,7 @@ import ldm.modules.midas as midas
 
 from ldm.util import instantiate_from_config
 
-from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization
+from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors
 from modules.paths import models_path
 from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting
 
@@ -333,8 +333,12 @@ def load_model(checkpoint_info=None):
 
     timer = Timer()
 
+    try:
         with sd_disable_initialization.DisableInitialization():
             sd_model = instantiate_from_config(sd_config.model)
+    except Exception as e:
+        print('Failed to create model quickly; will retry using slow method.', file=sys.stderr)
+        sd_model = instantiate_from_config(sd_config.model)
 
     elapsed_create = timer.elapsed()