• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

kubernetes-sigs / inference-perf / 40792758044
87%

Build:
DEFAULT BRANCH: main
Ran 29 Oct 2026 12:24AM UTC
Jobs 1
Files 137
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

29 Oct 2026 12:14AM UTC coverage: 85.662% (+0.1%) from 85.539%
40792758044

push

github

web-flow
Document server auto-detection as the default for model_name and tokenizer (#812)

Fixes #809

Old behavior: `server.model_name` and
`tokenizer.pretrained_model_name_or_path` had to be
set by hand. Leaving the model name unset took the first `/v1/models`
entry, but that
probe had no deadline and dropped the client certificate; pinning a
served alias gave the
tokenizer no way to reach the real model path; and SGLang/TGI dropped
`server.cert_path`/`key_path` before the client ever saw them.

New behavior: a config that sets neither is the documented default. Both
values are
detected from the server, the detection is bounded and authenticated, a
served alias falls
back to the entry's `root` -- with or without an explicit `model_name`
-- and report
generation counts tokens with the tokenizer the run actually resolved.

Changes:
- `openai_client.get_supported_models`: honor `load.request_timeout`
(bounded 30-second
fallback) and send `server.cert_path`/`key_path`, so an unresponsive or
mTLS server
  fails startup instead of hanging it or rejecting the probe.
- `SGlangModelServerClient`, `TGImodelServerClient`: accept and forward
the client
  certificate instead of dropping it.
- `openai_client`: an explicit `model_name` with no tokenizer path
probes `/v1/models`
only after the name itself fails to load, so a name that loads pays no
probe and no
error line; when the LoRA adapter check has already fetched the list,
the fallback
reuses it instead of sending a second request. The resolved tokenizer
config is
  exposed as `ModelServerClient.tokenizer_config`.
- `main.py` + `reportgen/base.py`: copy the resolved tokenizer config
into the run config
and build the report tokenizer through one helper that requires a path,
so a run that
auto-detects its tokenizer no longer reports without the streamed
token-timing
  correction.
- `docs/config.md`: document both fields as optional/auto-detected,
state when the root
fallback runs, and note that the resolved path lands... (continued)

60 of 61 new or added lines in 4 files covered. (98.36%)

1 existing line in 1 file now uncovered.

13215 of 15427 relevant lines covered (85.66%)

0.86 hits per line

Uncovered Changes

Lines Coverage ∆ File
1
17.84
1.67% inference_perf/main.py

Coverage Regressions

Lines Coverage ∆ File
1
79.17
-1.04% inference_perf/datagen/base.py
Jobs
ID Job ID Ran Files Coverage
1 40792758044.1 29 Oct 2026 12:24AM UTC 137
85.66
GitHub Action Run
Source Files on build 40792758044
  • Tree
  • List 137
  • Changed 8
  • Source Changed 6
  • Coverage Changed 8
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40792758044
  • 704b46e7 on github
  • Prev Build on main (#40729774242)
  • Next Build on main (#40807980347)
  • Delete
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc