[hbarclay]

Profiling counter totals in server_metrics_export.json (e.g. vllm:prompt_tokens, vllm:prefix_cache_queries) included the warmup phase: a PROFILING-tagged scrape near warmup start became the delta baseline, so the totals and timeslices started at warmup start even though phase_time_ranges reports the profiling window. Profiling samples are now limited to scrapes at or after the profiling start, so counters are deltas from the last scrape before profiling begins.

🤖 Generated with Claude Code

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@0ba8193fa812a3348a3b220232be9f6e6f3c4836

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@0ba8193fa812a3348a3b220232be9f6e6f3c4836

---

Last updated for commit: 0ba8193 • Browse code