Models converted via MLIR outputs only once

[kwon0408]

Environment (please complete the following information):

• OS/OS Version: Windows 11 24H2 build 26100.4652
• Source Version: master/2.19.0; master/2.19.0-p2
• Unity Version: Unity 6 (6000.0.27f1)

Describe the bug
If a TFLite model converted via MLIR is loaded, the model makes a successful output only for the first input. The second and further inputs are just ignored.

To Reproduce
I have uploaded the repro project to one of my repos.

The model is Fuel efficiency prediction from the TensorFlow tutorial.

Expected behavior
For each input the model should make an output.

Screenshots

Additional context
Inspecting the sample models with Netron, I found they are converted via TOCO (the old method).

[kwon0408]

I added a workaround at my repo, but I feel this is not the fundamental solution.

[asus4]

Hi @kwon0408, Thanks for reporting it.

I think I have resolved this use case at #402. Can you test it with the latest?

The input/output array was previously cached as GCHandle. In this PR, the GCHandle is updated when the array is changed.
In your case, you can also use the overloading method with NativeArray.