stream_options.include_usage not honored in chat completion streaming responses

[yaanfpv]

Summary

When a client requests stream_options.include_usage on a streaming chat completion, the final SSE chunk does not contain a usage field. This breaks benchmarking tools and observability pipelines that rely on per-request token accounting from the streaming response.

Affected versions

• vllm-swift v0.6.0
• vLLM 0.19.1

Reproduction

curl -N -s -X POST http://127.0.0.1:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<your model>",
    "messages": [{"role":"user","content":"Hello"}],
    "max_tokens": 20,
    "stream": true,
    "stream_options": {"include_usage": true}
  }'

Expected (per OpenAI spec): a final SSE data: chunk before data: [DONE] containing a usage object with prompt_tokens, completion_tokens, total_tokens.

Observed: the stream ends with data: [DONE] directly; no usage chunk is emitted. Clients that need streaming throughput numbers cannot recover them without re-tokenizing manually or making a non-streaming follow-up call.

Impact

• Benchmark scripts (e.g., a bench.py that tracks prefill / decode tok/s) cannot distinguish prompt vs completion tokens from the stream.
• Observability tools that aggregate per-request usage from streaming responses report zero or undefined.

A workaround is to scrape the engine's periodic loggers.py throughput line, but that is averaged and not per-request.

Reported by @yaanfpv.