PydanticUserError with defer_build=True and pydantic 2.12 in multiprocess forkserver workers

Bug Report

Summary

openai SDK 2.7.1 with pydantic 2.12.4 raises intermittent PydanticUserError ("Pydantic models should inherit from BaseModel, BaseModel cannot be instantiated directly") when deserializing chat completion responses in multiprocess forkserver worker processes.

Environment

• openai==2.7.1
• pydantic==2.12.4
• pydantic-core==2.41.5
• Python 3.12
• Using multiprocess (dill-based fork of multiprocessing) with forkserver start method
• Backend: vLLM serving client.chat.completions.create() responses

Reproduction

The error occurs when construct() calls _get_extra_fields_type(cls) which accesses cls.__pydantic_core_schema__["type"]. This triggers MockCoreSchema._get_built() → _attempt_rebuild() → model_rebuild(raise_errors=False, _parent_namespace_depth=5).

The hardcoded _parent_namespace_depth=5 walks 5 frames up the call stack to resolve forward references. When the call stack at that depth doesn't contain the openai types module namespace, model_rebuild silently fails and pydantic raises the error.

In a forkserver setup with ~700 workers processing millions of requests:

• With DEFER_PYDANTIC_BUILD=true (default): ~0.6% error rate, growing over time. The lazy rebuild retries on each request, mostly succeeding.
• With DEFER_PYDANTIC_BUILD=false: ~9.5% error rate. Eager build fails at import time for ~9.5% of forkserver workers, and since the failure is cached, ALL requests for those workers fail.

Traceback

openai/_models.py:231 construct()
→ openai/_models.py:405 _construct_field()
→ openai/_models.py:570 construct_type() (iterating choices list)
→ openai/_models.py:561 construct_type() → type_.construct(**value)
→ openai/_models.py:236 construct() → _get_extra_fields_type()
→ openai/_models.py:414 _get_extra_fields_type() → schema["type"] == "model"
→ pydantic/_internal/_mock_val_ser.py:41 __getitem__() → _get_built()
→ pydantic/_internal/_mock_val_ser.py:58 _get_built() → raises PydanticUserError

Related Issues

• #1306 (DEFER_PYDANTIC_BUILD causes issues)
• pydantic/pydantic#11849 (race condition with model instantiation)

Suggested Fix

Explicitly calling ChatCompletion.model_rebuild() (and related response types) before creating forkserver workers resolves the issue, since the correctly-built schema is then inherited by all worker processes. The SDK could do this internally — either eagerly at import time regardless of defer_build, or by catching the PydanticUserError in _get_extra_fields_type and calling model_rebuild() as a fallback.

Start by reproducing the forkserver failure with the reported Python, openai, pydantic, and multiprocess versions. Trace openai/_models.py through construct(), construct_type(), and _get_extra_fields_type(), then inspect pydantic/_internal/_mock_val_ser.py around _get_built(). Done means chat completion deserialization remains reliable across forkserver workers with both DEFER_PYDANTIC_BUILD settings, with regression coverage for the failure.