Magic header does not match the expected version.

[ChristophKarlHeck]

Hi guys,

I have integrated executorch into Mbed OS using an Arm Cortex M4 architecture (P-Nucleo-WB55RG). When I try to load the model, I get the following issue:

I executorch:main.cpp:68] Model PTE file loaded. Size: 960 bytes.
E executorch:program.cpp:108] Program identifier 'X��' != expected 'ET12'
I executorch:main.cpp:72] Program loading failed @ 0x0x20000E20: 0x23
F executorch:result.h:165] In function CheckOk(), assert failed: hasValue_

https://github.com/pytorch/executorch/blob/main/runtime/executor/program.cpp has caused the error in line 108.

How can I make sure that the magic header matches ET12?

Cheers, Christoph

[mergennachin]

@ChristophKarlHeck

This is related to backwards/forward compatibility with runtime and pte file. We are still in alpha mode; the ExecuTorch APIs and the .pte binary format may change in incompatible ways before stabilizing in a future beta release. When deploying models, we currently recommend using a version of the runtime built from the same git revision that was used to generate the .pte file.

See our disclaimer in our documentation landing page: https://pytorch.org/executorch/stable/index.html