VLLM2PCP(1)

📦 version 7.2.2, in arch rolling, centos-stream 10, fedora 43, fedora 44, ubuntu resolute, ubuntu stonking

💅 VLLM2PCP(1) — the vllmbench2pcp glow-up

okay bestie, so this is a page from the General Commands Manual, and it's literally all about vllmbench2pcp. sit down, grab a strawberry drink 🍓, we're doing this.

✨ NAME ✨

vllmbench2pcp — the vLLM benchmark results importer, no cap 🎀

✨ SYNOPSIS ✨

vllmbench2pcp [-hv] [-a archive] [-H hostname] results [...]

look at her. short, sweet, iconic 👑

✨ DESCRIPTION ✨

so vllmbench2pcp is a performance metrics importer tool, and what she does is create a PCP archive out of the JSON results of a vLLM benchmark. it's giving data transformation main character energy 🔥

what you get is an archive that spans the time for which the benchmark and its run iterations were active. and the way it works is: discrete PCP metrics get created for each recorded aspect of a run, and they get logged once at the start and once at the end of each iteration — followed by a mark record. periodt.

this means PCP tools can report on these metrics in a meaningful way alongside all the other system metrics (kernel, hardware, vLLM, and so on 🫶) that got gathered during the benchmark. slay.

the only mandatory command line argument is the JSON results file from a vllm bench invocation. now, this JSON can optionally be wrapped with additional Model Furnace metadata — and if that's detected, extra PCP labels get extracted and stored too (things like model and inference server names 👀). cute, right?

✨ OPTIONS ✨

the available command line options are, and yes we're listing all of them because nobody gets left out 💖:

  • -a archive, --archive=archive — specifies the name of the output PCP archive to be produced. by default, an archive name using the timestamp of the start of the benchmark run will be used. lowkey automatic, we love that for us.
  • -H hostname, --hostname=hostname — sets the hostname that will be stamped into the PCP archive label header. get her labeled 💅
  • -h, --help — display usage message and exit. the classic "what do I do" button, we stan.
  • -v, --verbose — report progress as each run is identified and its values saved. perfect for when you're nosy and want to watch everything happen in real time 🙈

✨ SEE ALSO ✨

go read these too, they're all in the same friend group:

bestie reminder: this page comes to you from Performance Co-Pilot / PCP, page VLLM2PCP(1). 💜 that's the whole vibe, we're done here, go make some archives ✨🌈
📦 where this page comes from