import ldm.modules.encoders.modules import open_clip import torch import transformers.utils.hub class DisableInitialization: """ When an object of this class enters a `with` block, it starts: - preventing torch's layer initialization functions from working - changes CLIP and OpenCLIP to not download model weights - changes CLIP to not make requests to check if there is a new version of a file you already have When it leaves the block, it reverts everything to how it was before. Use it like this: ``` with DisableInitialization(): do_things() ``` """ def __init__(self, disable_clip=True): self.replaced = [] self.disable_clip = disable_clip def replace(self, obj, field, func): original = getattr(obj, field, None) if original is None: return None self.replaced.append((obj, field, original)) setattr(obj, field, func) return original def __enter__(self): def do_nothing(*args, **kwargs): pass def create_model_and_transforms_without_pretrained(*args, pretrained=None, **kwargs): return self.create_model_and_transforms(*args, pretrained=None, **kwargs) def CLIPTextModel_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs): res = self.CLIPTextModel_from_pretrained(None, *model_args, config=pretrained_model_name_or_path, state_dict={}, **kwargs) res.name_or_path = pretrained_model_name_or_path return res 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' or url == 'openai/clip-vit-large-patch14' and args[0] == 'added_tokens.json': return None try: res = original(url, *args, local_files_only=True, **kwargs) if res is None: res = original(url, *args, local_files_only=False, **kwargs) return res except Exception: 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.replace(torch.nn.init, 'kaiming_uniform_', do_nothing) self.replace(torch.nn.init, '_no_grad_normal_', do_nothing) self.replace(torch.nn.init, '_no_grad_uniform_', do_nothing) if self.disable_clip: self.create_model_and_transforms = self.replace(open_clip, 'create_model_and_transforms', create_model_and_transforms_without_pretrained) self.CLIPTextModel_from_pretrained = self.replace(ldm.modules.encoders.modules.CLIPTextModel, 'from_pretrained', CLIPTextModel_from_pretrained) self.transformers_modeling_utils_load_pretrained_model = self.replace(transformers.modeling_utils.PreTrainedModel, '_load_pretrained_model', transformers_modeling_utils_load_pretrained_model) self.transformers_tokenization_utils_base_cached_file = self.replace(transformers.tokenization_utils_base, 'cached_file', transformers_tokenization_utils_base_cached_file) self.transformers_configuration_utils_cached_file = self.replace(transformers.configuration_utils, 'cached_file', transformers_configuration_utils_cached_file) self.transformers_utils_hub_get_from_cache = self.replace(transformers.utils.hub, 'get_from_cache', transformers_utils_hub_get_from_cache) def __exit__(self, exc_type, exc_val, exc_tb): for obj, field, original in self.replaced: setattr(obj, field, original) self.replaced.clear()