### What happened?

### Summary When using LiteLLM proxy with vLLM-backed models, enabling both `stream=True` and `logprobs=True` causes the stream to crash with a Pydantic v2 serialization error:

``` TypeError: 'MockValSer' object cannot be converted to 'SchemaSerializer' ```

This does not occur: - when `stream=False` - when `logprobs=False` - when using OpenAI-hosted models via LiteLLM

Direct calls to vLLM do support streaming + logprobs correctly, so the issue appears to be in LiteLLM’s streaming serialization logic.

- Crash happens after emitting the chunk, inside LiteLLM proxy: ``` File "litellm/proxy/proxy_server.py", line 4257 chunk = chunk.model_dump_json(exclude_none=True, exclude_unset=True)

pydantic_core.PydanticSerializationError: TypeError: 'MockValSer' object cannot be converted to 'SchemaSerializer' ```

### Expected Behavior

Streaming responses with `logprobs=True` should work the same for: - OpenAI models - vLLM-backed models Since vLLM itself supports streaming logprobs correctly, LiteLLM should be able to proxy them without crashing.

### Actual Behavior

- First streaming chunk arrives successfully - choice.logprobs.content is present and valid - LiteLLM c…