Python: Bug: When using Azure OpenAI, even if stream options are enabled, it is not reflected in the usage.

[yuichiromukaiyama]

Describe the bug

By setting stream_options.include_usage to true, the token usage is supposed to be returned, but it always ends up being None.

model: azure openai gpt 4o/4omini
api version: 2024-10-21

To Reproduce

async def stream_sample() -> None:
    kernel = sk.Kernel()
    service_id: str = "dummy"

    kernel.add_service(
        AzureChatCompletion(
            service_id=service_id,
            deployment_name=AZURE_OPENAI_COMPLETION_DEPLOYMENT_NAME,
            endpoint=AZURE_OPENAI_COMPLETION_ENDPOINT,
            api_key=AZURE_OPENAI_COMPLETION_API_KEY,
        )
    )

    service = kernel.get_service(service_id=service_id)
    settings = service.get_prompt_execution_settings_class()(service_id=service_id)

    if isinstance(settings, AzureChatPromptExecutionSettings):
        settings.extra_body = {
            "stream_options": {
                "include_usage": True,
            }
        }

    history = ChatHistory()
    history.add_user_message("hello")

    async for chunk in service.get_streaming_chat_message_contents(
        chat_history=history,
        settings=settings,
        kernel=kernel,
        arguments=KernelArguments(settings=settings),
    ):
        print("app chunk result: ", chunk)

[StreamingChatMessageContent(choice_index=0, inner_content=ChatCompletionChunk(id='chatcmpl-AVE9a9WOtYJbFwkUqcBrvMe6CWIRr', choices=[], created=1732005938, model='gpt-4o-mini', object='chat.completion.chunk', service_tier=None, system_fingerprint='fp_d54531d9eb', usage=None), ai_model_id='gpt-4o-mini', metadata={'id': 'chatcmpl-AVE9a9WOtYJbFwkUqcBrvMe6CWIRr', 'created': 1732005938, 'system_fingerprint': 'fp_d54531d9eb', 'usage': None, content_type='message', role=<AuthorRole.ASSISTANT: 'assistant'>, name=None, items=[], encoding=None, finish_reason=None)]

Expected behavior

[StreamingChatMessageContent(choice_index=0, inner_content=ChatCompletionChunk(id='chatcmpl-AVE9a9WOtYJbFwkUqcBrvMe6CWIRr', choices=[], created=1732005938, model='gpt-4o-mini', object='chat.completion.chunk', service_tier=None, system_fingerprint='fp_d54531d9eb', usage=CompletionUsage(completion_tokens=9, prompt_tokens=8, total_tokens=17)), ai_model_id='gpt-4o-mini', metadata={'id': 'chatcmpl-AVE9a9WOtYJbFwkUqcBrvMe6CWIRr', 'created': 1732005938, 'system_fingerprint': 'fp_d54531d9eb', 'usage': CompletionUsage(prompt_tokens=8, completion_tokens=9)}, content_type='message', role=<AuthorRole.ASSISTANT: 'assistant'>, name=None, items=[], encoding=None, finish_reason=None)]

Screenshots
none

Platform

• OS: Mac
• IDE: VS Code
• Language: Python
• Source: Semantic Kernel 1.15

Additional context
The cause is clear. The last chunk can sometimes have choice as 0, but it is being skipped entirely with continue. In the implementation of OpenAI, if usage is included, the chunk is processed appropriately.

azure

semantic-kernel/python/semantic_kernel/connectors/ai/open_ai/services/azure_chat_completion.py

Line 151 in 5197f0b

if len(chunk.choices) == 0:

openai

semantic-kernel/python/semantic_kernel/connectors/ai/open_ai/services/open_ai_chat_completion_base.py

Line 113 in 5197f0b

if len(chunk.choices) == 0 and chunk.usage is None:

[yuichiromukaiyama]

I was created pr: #9753