Add set_default_logger_callback Python API for custom ORT logging

[Copilot]

Description

Exposes a new onnxruntime.set_default_logger_callback(callback, severity=2) API that lets Python users route ORT log messages to any Python callable.

New function signature:

def set_default_logger_callback(
    callback: Callable[[int, str, str, str, str], None] | None,
    severity: int = 2,  # 0=Verbose, 1=Info, 2=Warning, 3=Error, 4=Fatal
) -> None: ...

Example:

import logging
import onnxruntime as ort

def ort_log_to_python(severity: int, category: str, logid: str, code_location: str, message: str) -> None:
    level = [logging.DEBUG, logging.INFO, logging.WARNING, logging.ERROR, logging.CRITICAL][severity]
    logging.log(level, "[ORT:%s] %s", category, message)

ort.set_default_logger_callback(ort_log_to_python, severity=0)  # capture all
sess = ort.InferenceSession("model.onnx")

ort.set_default_logger_callback(None)  # restore platform default (stderr/OutputDebugString)

Implementation:

• onnxruntime_pybind_state.cc: Adds PythonCallbackSink — a thread-safe ISink subclass that wraps a swappable Python callable and falls back to the platform sink when none is set. Callback exceptions are caught and swallowed (falling back to platform sink) to prevent crashes propagating into ORT. Adds set_default_logger_callback() in addGlobalMethods which updates the sink's callable in-place; no LoggingManager reconstruction needed.
• onnxruntime_pybind_module.cc: Modifies CreateOrtEnv() to replace the initial LoggingManager with one backed by PythonCallbackSink at module import time via a safe two-phase replace (destroy → recreate), valid because no sessions or threads exist yet.
• onnxruntime_pybind_module_functions.h: Declares CreateAndRegisterPythonCallbackSink() factory so the module init can call it.
• onnxruntime/__init__.py: Exports set_default_logger_callback in the public namespace.
• onnxruntime_test_python.py: Adds test_set_default_logger_callback covering normal use, None reset, inference stability, boundary severities, invalid-input rejection, and raising-callback safety.

Motivation and Context

The Python binding hard-codes nullptr as the ORT logging function at environment creation time, making it impossible to redirect ORT log output from Python. The C++ API already supports this via OrtApi::CreateEnvWithCustomLogger. This change brings equivalent capability to the Python API so users can integrate ORT logging with their application logger without resorting to file-descriptor manipulation.

[github-actions]

You can commit the suggested changes from lintrunner.

[JanSellner]

@tianleiwu Any news on the topic? :-)
The proposed API would be very useful if this would work!

[tianleiwu]

Review summary

The approach is sound — installing a PythonCallbackSink at module import and swapping the callable in-place (rather than rebuilding the LoggingManager singleton) is the right design, and the comments explaining the destroy → recreate sequence are excellent. Test coverage is good too.

There are two high-priority correctness/thread-safety issues around the lifecycle of the stored py::object that can affect all Python users (not just callers of the new API), plus a few suggestions. Details inline.

High-Priority

1. g_user_logging_callback defaults to a null handle, not py::none(). is_none() is ptr() == Py_None, which is false for a default-constructed (null) py::object. Since the sink is installed at import time, every ORT log at/above WARNING emitted before the first set_default_logger_callback() call takes the callback branch and invokes a null callable → null-pointer deref. Initialize the stored callback to py::none() (under the GIL) at sink creation.
2. Refcount manipulation without the GIL in SendImpl. ORT logs from internal worker threads that don't hold the GIL. py::object cb = g_user_logging_callback; does Py_INCREF before the GIL is acquired, and cb's destructor does Py_DECREF after the gil_scoped_acquire scope ends — both unsynchronized → data race. Hold the GIL around the copy → invoke → destroy sequence.

Suggestions

3. set_default_logger_callback(None) always rewrites the logger severity (default WARNING), silently undoing a prior set_default_logger_severity(0). Only update severity when a callback is provided, or document it.
4. err_msg in the exception handler is built but unused.
5. The callback is stored in a file-scope global rather than a sink member; consider a member for clarity.

Nitpicks

• onnxruntime_pybind_module_functions.h is missing a trailing newline.
• set_default_logger_callback import in __init__.py is out of alphabetical order (lint bot already flagged this block).

[Copilot]

Pull request overview

This PR adds a new Python API (onnxruntime.set_default_logger_callback) to allow Python callers to intercept ONNX Runtime log messages via a Python callable, integrating ORT logging with application logging without needing to rebuild sessions or manipulate file descriptors.

Changes:

• Add a PythonCallbackSink (ISink) implementation in the Python bindings and expose set_default_logger_callback(callback, severity=2).
• Replace the default LoggingManager during Python module initialization to route logs through the Python-aware sink.
• Export the new API from onnxruntime/__init__.py and add a Python test covering basic usage and invalid inputs.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
onnxruntime/python/onnxruntime_pybind_state.cc	Adds PythonCallbackSink and the set_default_logger_callback pybind entry point.
onnxruntime/python/onnxruntime_pybind_module.cc	Replaces the initial LoggingManager at import time to install the Python-aware sink.
onnxruntime/python/onnxruntime_pybind_module_functions.h	Declares the factory for creating/registering the Python callback sink.
onnxruntime/init.py	Re-exports set_default_logger_callback in the public Python namespace.
onnxruntime/test/python/onnxruntime_test_python.py	Adds a new unit test for the callback API surface and basic behavior.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.