Cannot load onnx checkpoint from huggingface repository #553
Comments
|
Thanks for the quick fix. It works for "multilingual-e5-base", however it fails for "multilingual-e5-large", which has a different directory structure: It now fails with a different error, though:
|
|
Oh yes this will be an issue for model for model >2GB, could you use instead : |
|
Unfortunately that fails
I was thinking something along the lines of @staticmethod
def _cached_file(
model_path: Union[Path, str],
use_auth_token: Optional[Union[bool, str]] = None,
revision: Optional[str] = None,
force_download: bool = False,
cache_dir: Optional[str] = None,
file_name: Optional[str] = None,
subfolder: str = "",
local_files_only: bool = False,
):
# locates a file in a local folder and repo, downloads and cache it if necessary.
model_path = Path(model_path)
if model_path.is_dir():
model_cache_path = model_path / file_name
else:
model_cache_path = snapshot_download(
repo_id=model_path.as_posix(),
allow_patterns=None if subfolder is None else subfolder+"/**",
use_auth_token=use_auth_token,
revision=revision,
cache_dir=cache_dir,
force_download=force_download,
local_files_only=local_files_only,
# local_dir="/tmp/openvinoconverter/",
# local_dir_use_symlinks="auto"
)
if subfolder is not None:
model_cache_path = os.path.join(model_cache_path, subfolder)
model_cache_path = os.path.join(model_cache_path, file_name)
model_cache_path = Path(model_cache_path)
return model_cache_pathmight solve the issue. At least it downloads all the right files. However, I suspect that there is another bug in openvino itself that prevents this from working unless I use this localdir="/tmp/openvinoconverter" line because of all the sym links. openvinotoolkit/openvino#22736 I'm not sure how to sensibly set this to a local directory, I'd much rather work only on the symlinks to the huggingface cache. |
|
The OpenVINO export for sentence transformers models should be fixed in optimum-intel v1.15.0, could you try : model = OVModelForFeatureExtraction.from_pretrained("intfloat/multilingual-e5-large", export=True) |
|
nice! now it works, thanks |
I am currently facing an issue while attempting to load an OVModelForFeatureExtraction from a pretrained checkpoint that is available as an ONNX file. The checkpoint is located in a subfolder within the Hugging Face repository.
I have tried to load the model using the following code snippet:
However, this approach fails with an error stating that
model.binis not available. It seems that the loading process does not consider the ONNX file (model.onnx) as it should. I have also attempted to explicitly set the file_name parameter as "model.onnx", but the issue persists.What works: Interestingly, when I clone the Hugging Face repository and load the model from the local onnx folder path, everything works correctly.
Expected Behavior: I expect the
from_pretrainedmethod to successfully load ONNX files from Hugging Face repositories, allowing me to directly use the pretrained checkpoint without needing to clone the repository locally.Traceback:
The text was updated successfully, but these errors were encountered: