Bug: `by_alias: NoneType` in openai-python v2.24.0 when logging.DEBUG is enabled

[GuillermoGaete]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

model_dump() in openai/_compat.py passes by_alias=None to pydantic's model_dump(), which causes pydantic-core's Rust serializer to raise TypeError: argument 'by_alias': 'NoneType' object cannot be converted to 'PyBool'.

This only triggers when the openai logger has level DEBUG enabled, because the affected model_dump() call is inside a if log.isEnabledFor(logging.DEBUG) block in _base_client.py:_build_request() (line 488).

The function signature in _compat.py:134 declares by_alias: bool | None = None, and passes it directly to pydantic without coercing to bool. Pydantic v2's Rust-based serializer does not accept None for by_alias.

To Reproduce

1. Enable DEBUG logging: logging.basicConfig(level=logging.DEBUG)
2. Make any API call via the openai client (sync or async)
3. The _build_request() method enters the debug logging block and calls model_dump(options, exclude_unset=True) without passing by_alias, so it defaults to None
4. pydantic-core raises: TypeError: argument 'by_alias': 'NoneType' object cannot be converted to 'PyBool'

Code snippets

import logging

# Enable DEBUG on the openai logger — this is the trigger
logging.basicConfig(level=logging.DEBUG)

from openai import AzureOpenAI  # also reproduces with OpenAI()

client = AzureOpenAI(
    azure_endpoint="https://example.openai.azure.com",
    api_key="fake-key",
    api_version="2025-01-01-preview",
)

# Any API call will crash before even reaching the network
client.chat.completions.create(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "hello"}],
)

# TypeError: argument 'by_alias': 'NoneType' object cannot be converted to 'PyBool'

OS

Linux (Debian-based Docker container, also reproducible on macOS)

Python version

Python v3.12.5

Library version

openai v2.24.0, pydantic v2.10.2, pydantic-core v2.27.1

[xXMrNidaXx]

I can reproduce this consistently with v2.24.0. The root cause is clear: in _compat.py, the model_dump() wrapper passes by_alias directly without coercing None to False, and pydantic v2's Rust serializer rejects None for that parameter.

Workaround until the fix lands:

import logging
# Simply keep your openai logger above DEBUG to avoid the affected code path
logging.getLogger('openai').setLevel(logging.INFO)

Or if you need DEBUG logging, you can monkey-patch the coercion:

import openai._compat as _compat
_original = _compat.model_dump

def _patched_model_dump(model, *, exclude_unset=False, by_alias=None, **kw):
    return _original(model, exclude_unset=exclude_unset, by_alias=bool(by_alias) if by_alias is not None else False, **kw)

_compat.model_dump = _patched_model_dump

Fix direction for maintainers: Change the function signature in _compat.py:134 from by_alias: bool | None = None to by_alias: bool = False, or add a by_alias = bool(by_alias) coercion before passing to pydantic. The pydantic-core Rust layer introduced strict type checking for this parameter.

[giulio-leone]

Fixed in #2940 — by_alias is now coerced to bool in model_dump() calls, preventing the NoneType error when logging.DEBUG is enabled.

[mvanhorn]

I've submitted a fix for this in PR #2953. Converts by_alias from None to False before passing to Pydantic v2's model_dump(), matching the pattern already used in the v1 fallback path.