Ep Context Model generated with external data is still dependent on the same data file

[BoarQing]

Describe the issue

We found when the Ep Context Model is generated from a model with an external data, if an OP is not at the provider domain, it would still read to the original data file. This is undesirable since the external data is quite big.

Maybe MSFT can trim the external data for the new Ep Context Model to only contains those OPs' weights?
Or Maybe MSFT can just embed those weights into the new Ep Context Model?

To reproduce

1. Generate an Ep context model on a model with external data.
2. Delete the external data.
3. Inferencing the Ep context model would crash for not finding the external data.

Urgency

It is important for MSFT release.

Platform

Windows

OS Version

Windows11

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

main

ONNX Runtime API

C++

Architecture

X64

Execution Provider

Vitis AI

Execution Provider Library Version

No response

[BoarQing]

@snnn @jywu-msft @HectorSVC

[BoarQing]

#23374 A temporary solution to move all the tensor into the .onnx.

[HectorSVC]

Please give it a try with latest main.