---

This section details on the original issue you should resolve

<issue_title>[CodeDesignBug] Inline IsOnnxStaticRegistrationDisabled() breaks schema registration in external modules</issue_title>
<issue_description># Bug Report

Is the issue related to model conversion?

No, this issue is not related to model conversion. It is a design problem in ONNX core related to how static registration is exposed through the function IsOnnxStaticRegistrationDisabled().

Describe the bug

IsOnnxStaticRegistrationDisabled() is defined as an inline function in onnx/defs/operator_sets.h:

inline bool IsOnnxStaticRegistrationDisabled() {
#ifdef __ONNX_DISABLE_STATIC_REGISTRATION
  return true;
#else
  return false;
#endif
}

This design causes a packaging and consistency problem because its behavior depends on the compile-time macro __ONNX_DISABLE_STATIC_REGISTRATION. If another module (e.g., ONNX Runtime or a user application) calls this function, it must define the same macro at compile time to get a consistent result.
Otherwise, this breaks encapsulation and makes it difficult to use ONNX as a prebuilt library or in mixed build environments.

System information

ONNX 1.19.1

Reproduction instructions

1. Build ONNX with __ONNX_DISABLE_STATIC_REGISTRATION defined.
2. In a separate project (e.g., ONNX Runtime or your own application), include ONNX headers without defining this macro.
3. Call bool disabled = IsOnnxStaticRegistrationDisabled();
4. Observe that the return value does not reflect the actual ONNX build configuration.
   This can cause inconsistent behavior, such as conditional schema registration being executed incorrectly.

Expected behavior

IsOnnxStaticRegistrationDisabled() should always reflect the actual build-time configuration of the ONNX library, regardless of the caller’s compile-time macros.

My suggestion

To achieve this, the function should not be inlined in the header.
Instead:

• Only declare the prototype in the header.
• Implement it in a .cc file (e.g., schema.cc) where it can access the real build-time flag or internal variable.
  This ensures consistent behavior across all binaries linking against ONNX, improving encapsulation and packaging reliability.
• Since this function is only used for schema registration, there is no need for it to be inlined.

Notes

ONNX Runtime needs to call

RegisterOnnxOperatorSetSchema()
RegisterOnnxMLOperatorSetSchema()
RegisterOnnxTrainingOperatorSetSchema()

These calls are necessary only when ONNX static registration is disabled.
If static registration is already enabled, invoking them again will cause duplicate schema registration errors.

Since this function’s return value depends on the compile-time macro __ONNX_DISABLE_STATIC_REGISTRATION, external projects (like ONNX Runtime) must manually ensure that their build flags match ONNX’s internal configuration — which breaks encapsulation and can easily lead to inconsistent behavior.

Additionally, when static registration is enabled, some registration functions such as RegisterOnnxTrainingOperatorSetSchema() and RegisterOnnxPreviewOperatorSetSchema() are always called inside SchemasRegisterer(). It allocates memory for all training/preview schemas, even if some of them are not used by the model. This may lead to unnecessary memory usage, particularly on memory-constrained devices.</issue_description>

Comments on the Issue (you are @copilot in this section)