onnxruntime 1.24 broke loading models from the huggingface_hub cache

[istupakov]

Describe the issue

HuggingFace is the most popular method for distributing models. When downloaded through the official huggingface_hub package, it saves them to a cache. The snapshot_download function returns a folder with symlinks to the actual blobs.

onnxruntime 1.24 prohibits (#26776) the use of symlinks for data files, which completely breaks the loading of large models from the huggingface_hub cache.

To reproduce

Code

!uv pip install huggingface_hub "onnxruntime==1.24.1"

from pathlib import Path
from huggingface_hub import snapshot_download
import onnxruntime as rt

model_folder = snapshot_download("istupakov/parakeet-tdt-0.6b-v3-onnx")
rt.InferenceSession(Path(model_folder) / "encoder-model.onnx")

Gives an error:

Fail: [ONNXRuntimeError] : 1 : FAIL : External data path validation failed for initializer: pre_encode.out.bias. Error: tensorprotoutils.cc:346 ValidateExternalDataPath External data path: "encoder-model.onnx.data" escapes model directory: "/root/.cache/huggingface/hub/models--istupakov--parakeet-tdt-0.6b-v3-onnx/snapshots/abd2878d52a678ce380088ef9d9b1d9664404565"

Urgency

This issue completely broke my onnx-asr project (and probably many others using huggingface_hub) running on onnxruntime 1.24.

Platform

Linux

OS Version

Any

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.24.1

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[istupakov]

Will likely be fixed in #27352. I really hope it makes it into 1.24.2!