fix(types): call model_rebuild() on openai SDK streaming types to fix MockValSer crash

openai._models.BaseModel sets defer_build=True, so SDK subclasses start with MockValSer
as their pydantic serializer. model_validate() — used by the SDK for all response parsing
— does not trigger the deferred build (only __init__ does). So streaming objects like
ChoiceLogprobs and ChoiceDeltaToolCall retain MockValSer after parsing.

streaming_handler.py stores these raw SDK objects as extra field values in
ModelResponseStream. When proxy_server._serialize_streaming_chunk() calls model_dump_json(),
pydantic's Rust core looks up type(value).__pydantic_serializer__ at the C level, finds
MockValSer, and raises TypeError.

model_rebuild() at module load time forces the deferred build to complete, installing a
real SchemaSerializer before any request is processed.

Fixes: #24357, #18801, #30617

litellm/types/utils.py

@@ -3791,3 +3791,53 @@ class GenericGuardrailAPIInputs(TypedDict, total=False):
         AllMessageValues
     ]  # structured messages sent to the LLM - indicates if text is from system or user
     model: Optional[str]  # the model being used for the LLM call
+
+
+# openai._models.BaseModel sets defer_build=True, so subclasses start with MockValSer
+# as their __pydantic_serializer__. model_validate() (used by the SDK for all response
+# parsing) does NOT trigger the deferred build — only __init__ does. So every SDK object
+# produced by parsing a streaming chunk has MockValSer at the class level.
+#
+# streaming_handler.py stores these raw SDK objects as extra='allow' field values in
+# litellm's ModelResponseStream. When proxy_server calls model_dump_json(), the Rust
+# SchemaSerializer looks up type(value).__pydantic_serializer__ at the C level (bypassing
+# Python's __getattr__ retry), finds MockValSer, and raises:
+#   TypeError: 'MockValSer' object cannot be converted to 'SchemaSerializer'
+#
+# The crash is intermittent because model_dump() (Python path) accidentally self-heals
+# via MockValSer.__getattr__ -> attempt_rebuild(). If model_dump() runs first (usage
+# chunk), the SDK class is rebuilt as a side effect and model_dump_json() works. If
+# model_dump_json() runs first, it crashes.
+#
+# Remove these calls only when streaming_handler stops storing raw SDK objects in litellm
+# models (i.e., always converts to dicts or litellm's own types before assignment).
+from openai.types.chat import ChatCompletionChunk as _ChatCompletionChunk
+from openai.types.chat.chat_completion_chunk import (
+    Choice as _OAIChoice,
+    ChoiceDelta as _OAIChoiceDelta,
+    ChoiceLogprobs as _OAIChoiceLogprobs,
+    ChoiceDeltaToolCall as _OAIChoiceDeltaToolCall,
+    ChoiceDeltaToolCallFunction as _OAIChoiceDeltaToolCallFunction,
+)
+from openai.types.chat.chat_completion_token_logprob import (
+    ChatCompletionTokenLogprob as _OAIChatCompletionTokenLogprob,
+    TopLogprob as _OAITopLogprob,
+)
+
+
+def _rebuild_sdk_streaming_types() -> None:
+    for cls in (
+        _ChatCompletionChunk,
+        _OAIChoice,
+        _OAIChoiceDelta,
+        _OAIChoiceLogprobs,
+        _OAIChoiceDeltaToolCall,
+        _OAIChoiceDeltaToolCallFunction,
+        _OAIChatCompletionTokenLogprob,
+        _OAITopLogprob,
+    ):
+        cls.model_rebuild()
+
+
+_rebuild_sdk_streaming_types()
+del _rebuild_sdk_streaming_types

tests/test_litellm/types/test_types_utils.py

@@ -344,6 +344,80 @@ def test_parallel_request_limiter_internal_fields_in_all_litellm_params():
         )
 
 
+def test_sdk_streaming_types_model_dump_json_after_model_validate():
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/18801 and #24357.
+
+    SDK objects created via model_validate() carry MockValSer as their pydantic serializer.
+    When stored as extra fields in ModelResponseStream and serialized via model_dump_json(),
+    pydantic's Rust core hits MockValSer at the C level and raises TypeError.
+    """
+    from openai.types.chat.chat_completion_chunk import (
+        ChoiceDeltaToolCall as _OAIChoiceDeltaToolCall,
+        ChoiceLogprobs as _OAIChoiceLogprobs,
+    )
+    from openai.types.chat.chat_completion_token_logprob import (
+        ChatCompletionTokenLogprob as _OAIChatCompletionTokenLogprob,
+    )
+
+    from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
+
+    logprobs_obj = _OAIChoiceLogprobs.model_validate(
+        {
+            "content": [
+                _OAIChatCompletionTokenLogprob.model_validate(
+                    {"token": "hello", "bytes": [104], "logprob": -0.5, "top_logprobs": []}
+                )
+            ]
+        }
+    )
+    tool_call_obj = _OAIChoiceDeltaToolCall.model_validate(
+        {"index": 0, "id": "call_abc", "type": "function", "function": {"name": "my_tool", "arguments": '{"x": 1}'}},
+    )
+
+    model_response = ModelResponseStream(
+        id="chatcmpl-test",
+        choices=[StreamingChoices(finish_reason=None, index=0, delta=Delta(role="assistant"), logprobs=logprobs_obj)],
+        model="gpt-4o",
+    )
+    model_response.choices[0].delta.tool_calls = [tool_call_obj]
+
+    result = model_response.model_dump_json(exclude_none=True, exclude_unset=True)
+    assert "my_tool" in result
+    assert "hello" in result
+
+
+def test_sdk_streaming_empty_logprobs_model_dump_json():
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/24357.
+
+    The final stop chunk often carries ChoiceLogprobs(content=None, refusal=None).
+    This empty logprobs object is still a raw SDK type created via model_validate()
+    and retains MockValSer, causing model_dump_json() to crash the same way.
+    """
+    from openai.types.chat.chat_completion_chunk import ChoiceLogprobs as _OAIChoiceLogprobs
+
+    from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
+
+    empty_logprobs = _OAIChoiceLogprobs.model_validate({"content": None, "refusal": None})
+
+    model_response = ModelResponseStream(
+        id="chatcmpl-test",
+        choices=[
+            StreamingChoices(
+                finish_reason="stop",
+                index=0,
+                delta=Delta(role="assistant", content=None),
+                logprobs=empty_logprobs,
+            )
+        ],
+        model="gpt-4o",
+    )
+
+    result = model_response.model_dump_json(exclude_none=True, exclude_unset=True)
+    assert "stop" in result
+
+
 def test_delta_maps_reasoning_to_reasoning_content():
     """
     Test that Delta maps 'reasoning' field to 'reasoning_content'.