The runtime images - both onnx-eis and onnx-runtime - assume that the downloaded image (from OCI or HF) is an onnx model, and that the cache in another location is where the .bundle files are. This works well if your process always is to download onnx and then compile locally.

But a compiled .bundle may be pushed to a registry, and may already exist there. Someone may download a .bundle directly, which means that will be in the HF cache.

We need to be able to run the .bundle models directly from the HF cache, not just the local compiled versions.