[Bug]: Loading models with additional files fails with onnxruntime 1.24.1

[NikolasWolke]

What happened?

Loading either intfloat/multilingual-e5-large or jinaai/jina-embeddings-v3 fails with onnxruntime==1.24.1.

To reproduce create an instance of the model, e.g. TextEmbedding(model_name="jinaai/jina-embeddings-v3"). The results in FAIL : External data path validation failed for initializer […] Error: tensorprotoutils.cc:346 ValidateExternalDataPath External data path: "model.onnx_data" escapes model directory.

This is most likely due to a new security check that was added in that release.

I suspect, any model that specifies additional_files in the BaseModelDescription is affected, but have only tested the two mentioned above. Manually setting the cache_dir does not change the behavior.

Environment

python 3.14.2
fastembed==0.7.4
onnxruntime==1.24.1

What is the expected behaviour?

The model should load and be ready for inference.

A minimal reproducible example

TextEmbedding(model_name="jinaai/jina-embeddings-v3")

What Python version are you on? e.g. python --version

Python 3.14.2

FastEmbed version

v0.7.4

What os are you seeing the problem on?

Linux

Relevant stack traces and/or logs

Traceback (most recent call last):
    TextEmbedding(model_name="jinaai/jina-embeddings-v3")
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "../.venv/lib/python3.14/site-packages/fastembed/text/text_embedding.py", line 114, in __init__
    self.model = EMBEDDING_MODEL_TYPE(
                 ~~~~~~~~~~~~~~~~~~~~^
        model_name=model_name,
        ^^^^^^^^^^^^^^^^^^^^^^
    ...<6 lines>...
        **kwargs,
        ^^^^^^^^^
    )
    ^
  File "../.venv/lib/python3.14/site-packages/fastembed/text/multitask_embedding.py", line 49, in __init__
    super().__init__(*args, **kwargs)
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "../.venv/lib/python3.14/site-packages/fastembed/text/onnx_embedding.py", line 259, in __init__
    self.load_onnx_model()
    ~~~~~~~~~~~~~~~~~~~~^^
  File "../.venv/lib/python3.14/site-packages/fastembed/text/onnx_embedding.py", line 324, in load_onnx_model
    self._load_onnx_model(
    ~~~~~~~~~~~~~~~~~~~~~^
        model_dir=self._model_dir,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
    ...<5 lines>...
        extra_session_options=self._extra_session_options,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "../.venv/lib/python3.14/site-packages/fastembed/text/onnx_text_model.py", line 59, in _load_onnx_model
    super()._load_onnx_model(
    ~~~~~~~~~~~~~~~~~~~~~~~~^
        model_dir=model_dir,
        ^^^^^^^^^^^^^^^^^^^^
    ...<5 lines>...
        extra_session_options=extra_session_options,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "../.venv/lib/python3.14/site-packages/fastembed/common/onnx_model.py", line 108, in _load_onnx_model
    self.model = ort.InferenceSession(
                 ~~~~~~~~~~~~~~~~~~~~^
        str(model_path), providers=onnx_providers, sess_options=so
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "../.venv/lib/python3.14/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 504, in __init__
    self._create_inference_session(providers, provider_options, disabled_optimizers)
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "../.venv/lib/python3.14/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 610, in _create_inference_session
    sess.initialize_session(providers, provider_options, disabled_optimizers)
    ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : External data path validation failed for initializer: roberta.embeddings.word_embeddings.parametrizations.weight.0.lora_B. Error: tensorprotoutils.cc:346 ValidateExternalDataPath External data path: "model.onnx_data" escapes model directory: "/tmp/fastembed_cache/models--jinaai--jina-embeddings-v3/snapshots/f1944de8402dcd5f2b03f822a4bc22a7f2de2eb9/onnx"

[joein]

Hey @NikolasWolke

Thanks for reporting it!
We'll address onnxruntime 1.24 issues soon! (Hopefully next week)

[cau-git]

This appears to be happening universally with HuggingFace models which download to the HF cache (~/.cache/huggingface), because the model versioning is realized with symlinks that reference to blobs outside the model's base directory. It would be great if this can be relaxed or optionally disabled on onnxruntime.

[yuslepukhin]

We will take a look at it. We will need to collect some requirements from other stakeholders.

[yuslepukhin]

Would creating an option to specify a list of good folders be a good solution?