InvokeAI
InvokeAI copied to clipboard
invoke-ai
[bug]: Invoke doesn't start with some embeddings
Is there an existing issue for this?
- [X] I have searched the existing issues
OS
Linux
GPU
cuda
VRAM
No response
What happened?
When launching with some embeddings (I believe they were either trained https://github.com/AUTOMATIC1111/stable-diffusion-webui or https://github.com/cafeai/stable-textual-inversion-cafe) Invoke just crashes with
╭────────────────────────── Traceback (most recent call last) ───────────────────────────╮
│ /home/ashtaka/AI/InvokeAI/scripts/invoke.py:4 in <module> │
│ │
│ 1 #!/usr/bin/env python │
│ 2 │
│ 3 import ldm.invoke.CLI │
│ ❱ 4 ldm.invoke.CLI.main() │
│ 5 │
│ 6 │
│ │
│ /home/ashtaka/AI/InvokeAI/ldm/invoke/CLI.py:117 in main │
│ │
│ 114 │ │
│ 115 │ # preload the model │
│ 116 │ try: │
│ ❱ 117 │ │ gen.load_model() │
│ 118 │ except AssertionError: │
│ 119 │ │ emergency_model_reconfigure() │
│ 120 │ │ sys.exit(-1) │
│ │
│ /home/ashtaka/AI/InvokeAI/ldm/generate.py:818 in load_model │
│ │
│ 815 │ │ ''' │
│ 816 │ │ preload model identified in self.model_name │
│ 817 │ │ ''' │
│ ❱ 818 │ │ self.set_model(self.model_name) │
│ 819 │ │
│ 820 │ def set_model(self,model_name): │
│ 821 │ │ """ │
│ │
│ /home/ashtaka/AI/InvokeAI/ldm/generate.py:858 in set_model │
│ │
│ 855 │ │ │
│ 856 │ │ seed_everything(random.randrange(0, np.iinfo(np.uint32).max)) │
│ 857 │ │ if self.embedding_path is not None: │
│ ❱ 858 │ │ │ self.model.embedding_manager.load( │
│ 859 │ │ │ │ self.embedding_path, self.precision == 'float32' or self.precis │
│ 860 │ │ │ ) │
│ 861 │
│ │
│ /home/ashtaka/AI/InvokeAI/ldm/modules/embedding_manager.py:238 in load │
│ │
│ 235 │ │ │ ckpt_paths = [ckpt_paths] │
│ 236 │ │ ckpt_paths = self._expand_directories(ckpt_paths) │
│ 237 │ │ for c in ckpt_paths: │
│ ❱ 238 │ │ │ self._load(c,full) │
│ 239 │ │ # remember that we know this term and don't try to download it again fro │
│ 240 │ │ # note that if the concept name is also provided and different from the │
│ 241 │ │ # both will be stored in this dictionary │
│ │
│ /home/ashtaka/AI/InvokeAI/ldm/modules/embedding_manager.py:276 in _load │
│ │
│ 273 │ │ │ │ print(f">> {ckpt_path} has >1 embedding, only the first will be │
│ 274 │ │ │ │
│ 275 │ │ │ string_to_param_dict = ckpt['string_to_param'] │
│ ❱ 276 │ │ │ embedding = list(string_to_param_dict.values())[0] │
│ 277 │ │ │ self.add_embedding(token_str, embedding, full) │
│ 278 │ │ │
│ 279 │ │ # Handle .bin textual inversion files from Huggingface Concepts │
│ │
│ /home/ashtaka/.conda/envs/invokeai/lib/python3.9/site-packages/torch/nn/modules/contai │
│ ner.py:714 in values │
│ │
│ 711 │ def values(self) -> Iterable[Any]: │
│ 712 │ │ r"""Return an iterable of the ParameterDict values. │
│ 713 │ │ """ │
│ ❱ 714 │ │ return (self[k] for k in self._keys) │
│ 715 │ │
│ 716 │ def update(self, parameters: Union[Mapping[str, Any], 'ParameterDict']) -> N │
│ 717 │ │ r"""Update the :class:`~torch.nn.ParameterDict` with the key-value pairs │
│ │
│ /home/ashtaka/.conda/envs/invokeai/lib/python3.9/site-packages/torch/nn/modules/module │
│ .py:1207 in __getattr__ │
│ │
│ 1204 │ │ │ modules = self.__dict__['_modules'] │
│ 1205 │ │ │ if name in modules: │
│ 1206 │ │ │ │ return modules[name] │
│ ❱ 1207 │ │ raise AttributeError("'{}' object has no attribute '{}'".format( │
│ 1208 │ │ │ type(self).__name__, name)) │
│ 1209 │ │
│ 1210 │ def __setattr__(self, name: str, value: Union[Tensor, 'Module']) -> None: │
╰────────────────────────────────────────────────────────────────────────────────────────╯
AttributeError: 'ParameterDict' object has no attribute '_keys'
Screenshots
There is nothing to screenshot
Additional context
Here is one of the embeddings that crashes https://cdn.discordapp.com/attachments/1049117806605643846/1049143254337196042/crashfever-style-34000.pt
And here is a discord thread https://discord.com/channels/1020123559063990373/1049117806605643846/1049117806605643846
Contact Details
discord: Ashtaka#0043
other embeddings have the embedding data in string_to_param_dict[0], ones like this somehow decide to make string_to_param_dict not a dict and just drop the embedding right here. this needs to be handled by inspecting the type at runtime.
a problem we have with these files is that they are all called .pt and that extension says nothing other than file saved by pytorch, so figuring out what they are is a case of setting a breakpoint and dumping out their contents to see if they can be interpreted somehow. so even if we get one version working there's no guarantee some other version isn't going to just pop up. it's like the xkcd standards comic but there's no standards.
Somehow Auto1111 deals with it fine :( But their code is a mess… So how could I make these embeddings work properly with Invoke?
I have the same issue but mine ends with Pickle exhausted before seeing STOP
for more info, I added the embeddings from huggingface into the embedding folder within invokeai and it stops opening until I remove a specific amount of them. im assuming its a particular file but the cmd closes before I can see it. If I do manage to get some loaded it then says certain prompts are already loaded and basically takes 10+ minutes to change a model
Dont know if im just being dumb but now I get this with the embeddings
really no idea what's wrong here, sorry, i don't have enough information to give you any kind of answer at all.
Oh sorry, when I add the embedding from huggingface, I get this error when opening invoke and it crashes.
🤔 hmm.
what is the filename of the embedding?
fwiw - for embeddings in the huggingface concepts library you can just prompt <embedding-name> and Invoke will download it for you automatically. best to try this out on the CLI version because then you get feedback about whether it's working.
I may be an idiot but I downloaded all 800 or so of them. Haha but I didn't realize I could do that!