Count reasoning_content in the restful chat benchmark

[David-Wu1119]

Motivation

With --backend lmdeploy-chat, benchmark/profile_restful_api.py only treats delta.content as output. A server started with --reasoning-parser streams the model's thinking as delta.reasoning_content, so:

• TTFT is measured to the first answer token after all of the reasoning, or stays 0.0 when the whole max_tokens budget goes to reasoning (the request still counts as successful, so the 0 is averaged in);
• ITL drops every reasoning token, and the retokenized output count only covers the answer.

Against a fake OpenAI-compatible server streaming 64 tokens per request (first token after 100 ms, then one every 10 ms), 20 prompts:

reasoning tokens of 64		Mean TTFT (ms)	Mean TPOT (ms)	Mean ITL (ms)	retokenized output
64	before	0.00	13.57	0.00	0 of 1280
64	after	111.60	11.79	11.79	1280 of 1280
40	before	586.75	4.39	12.04	480 of 1280
40	after	106.55	12.07	12.07	1280 of 1280

lmdeploy's own benchmark/benchmark_chat_completion.py already counts content or reasoning_content as the first token, and sglang's bench_serving.py (which this script follows) counts reasoning_content as content too.

Modification

In async_request_openai_chat_completions, count reasoning_content + content of each delta as output, reading both with or '' since either can be null. The same lines also stop assuming choices is non-empty, so a usage-only chunk with choices: [] no longer raises IndexError.

No open issue or PR covers this (searched for reasoning_content, profile_restful_api reasoning, TTFT reasoning, reasoning-parser benchmark).

BC-breaking (Optional)

No. Non-reasoning output is measured as before.

Checklist

1. pre-commit (ruff, docformatter, the copyright hook and the other configured hooks) passes on the changed file.
2. There are no unit tests for the benchmark scripts; the table above comes from the fake-server runs.

Written with AI assistance (Claude Code).

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.