gpu-index-26.1.0.0.0 Private AI container crashes on startup — ONNX model load failure with GPU mode

Question

I'm running the Oracle Private AI Services Container gpu-index-26.1.0.0.0 (build 2.1.2) on Oracle Linux with Podman, on a host with multiple NVIDIA GPUs (driver 610.57.04, CUDA 13.1). nvidia-smi and nvidia-ctk cdi list both confirm the GPUs and driver are healthy and correctly exposed to Podman via --gpus all / CDI.

The container starts, but exits a few seconds later. The CPU-only image (private-ai:25.1.3.0.0) using the identical model/config/security directories starts and runs without issue, which rules out a bad model file or config problem.

Container log shows:

Error starting Micronaut server: Bean definition [com.oracle.pai.config.EagerLoaderHelper] could not be loaded
Message: Failed to load the onnx model file EM_ALL_MINILM_L12_V2.onnx with ONNX Runtime version 1.20.1.
Caused by: com.oracle.oaa.scoring.exception.InputFormatException: Failed to load the onnx model file EM_ALL_MINILM_L12_V2.onnx with ONNX Runtime version 1.20.1.

No further native error detail is surfaced in the logs. Internal config shows PRIVATE_AI_GPU_MODE : FORCE, so the app requires GPU-based ONNX inference and doesn't fall back to CPU.

Has anyone else hit this on the same image version? Is there a known additional package/dependency required for GPU ONNX inference that isn't covered in the current docs/FAQ?

Gianni Ceresa · Answer

Hi,

Welcome to the Oracle AI and Analytics community.

Your question is currently in the AIDP category, which isn't the right place.

"Technically" the Oracle Private AI Services container is a database feature and therefore should be covered in forums covering the database. But I'm not fully sure which one would be the right one. It's a quite recent topic, not having had much coverage in forums so far.

Let me see if I can figure out what community is supposed to cover it, so you can get the best audience for your need. In the meantime will move the question to the AI Services category (still not the right place as this isn't an OCI AI Service, but better than AIDP).