Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

When an upstream server or proxy callback injects custom attributes into SSE streaming chunks, preserve_upstream_non_openai_attributes() correctly copies them to the ModelResponseStream for content-bearing chunks — but not for the final chunk (the one with finish_reason set and empty/null content).

This means any custom metadata attached to the last SSE chunk by server-side callbacks or middleware is silently dropped by the SDK. The client never sees it.

Root cause

In litellm/litellm_core_utils/streaming_handler.py, method return_processed_chunk_logic():

• Content-bearing chunks (~line 917): preserve_upstream_non_openai_attributes() is called — custom attributes are preserved ✅
• Final chunk (elif self.received_finish_reason is not None, ~line 948): this code path returns model_response without calling preserve_upstream_non_openai_attributes() — custom attributes are dropped ❌

The fix should be a single call to preserve_upstream_non_openai_attributes(model_response, original_chunk) in the received_finish_reason branch, before return model_response at line 981.

Steps to Reproduce

1. Set up a proxy callback that injects a custom field into the last SSE chunk:

async def async_post_call_streaming_iterator_hook(self, user_api_key_dict, response, request_data):
    async for chunk in response:
        if chunk.choices and chunk.choices[0].finish_reason:
            chunk.custom_field = {"key": "value"}
        yield chunk

1. On the client, consume the stream and check for the custom field:

response = litellm.completion(model="gpt-4", messages=[...], stream=True)
for chunk in response:
    data = chunk.model_dump()
    if "custom_field" in data:
        print("Found custom_field")  # Never prints

1. The same request via raw HTTP (parsing SSE lines directly) does show custom_field in the last chunk's JSON.

Expected behavior

preserve_upstream_non_openai_attributes() should be called for all chunks that are returned to the client, including the final chunk with finish_reason. The function already exists and works correctly — it just needs to be called in the received_finish_reason branch.

What part of LiteLLM is this about?

SDK, Proxy

What LiteLLM version are you on?

v1.80.10 (also verified on v1.82.x — same code path)