error when loading custom VAE from safetensors file

[ssube]

When converting a model that uses a custom VAE, and the vae key is set to a local file rather than a HF repo, it fails with:

[2023-05-27 17:04:19,193] INFO: 299372 139862552576000 onnx_web.convert.diffusion.diffusers: loading custom VAE: ../models/.cache/vae-ft-mse-840000-ema-pruned.safetensors
loading pretrained ../models/.cache/vae-ft-mse-840000-ema-pruned.safetensors True False                                                                                                                                                     [2023-05-27 17:04:19,255] ERROR: 299372 139862552576000 __main__: error converting diffusion model diffusion-fantastic-mix-25d-v2
Traceback (most recent call last):                                                                                                                                                                                                          
  File "/opt/onnx-web/api/onnx_env/lib/python3.10/site-packages/diffusers/configuration_utils.py", line 410, in load_config
    config_dict = cls._dict_from_json_file(config_file)                                                                                                                                                                                     
  File "/opt/onnx-web/api/onnx_env/lib/python3.10/site-packages/diffusers/configuration_utils.py", line 520, in _dict_from_json_file                                                                                                            text = reader.read()
  File "/usr/lib/python3.10/codecs.py", line 322, in decode                                                                                                                                                                                 
    (result, consumed) = self._buffer_decode(data, self.errors, final)                                         
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x8f in position 25730: invalid start byte                                                                                                                                              
                                                                                                                                                                                                                                            
During handling of the above exception, another exception occurred:                                                                                                                                                                         
                                                                                                                      
Traceback (most recent call last):                                                                                                                                                                                                          
  File "/opt/onnx-web/api/onnx_web/convert/__main__.py", line 359, in convert_models                                                                                                                                                        
    converted, dest = convert_diffusion_diffusers(                                                                                                                                                                                          
  File "/opt/onnx-web/api/onnx_env/lib/python3.10/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "/opt/onnx-web/api/onnx_web/convert/diffusion/diffusers.py", line 525, in convert_diffusion_diffusers
    vae = AutoencoderKL.from_pretrained(replace_vae)
  File "/opt/onnx-web/api/onnx_env/lib/python3.10/site-packages/diffusers/models/modeling_utils.py", line 494, in from_pretrained
    config, unused_kwargs, commit_hash = cls.load_config(
  File "/opt/onnx-web/api/onnx_env/lib/python3.10/site-packages/diffusers/configuration_utils.py", line 414, in load_config
    raise EnvironmentError(f"It looks like the config file at '{config_file}' is not a valid JSON file.")
OSError: It looks like the config file at '../models/.cache/vae-ft-mse-840000-ema-pruned.safetensors' is not a valid JSON file.

The error is technically correct, the safetensors file is not a JSON file. For some reason, AutoencoderKL.from_pretrained is now looking for a config at the same filename/path, which does not exist and causes the error. I'm not sure how this worked before, but it must have been loading a VAE config from the working Torch model or something.

[ssube]

This has been updated to use AutoencoderKL.from_single_file and should support safetensors and pickle tensors.