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

kubernetes-sigs / inference-perf / 40807980347
87%

Build:
DEFAULT BRANCH: main
Ran 29 Oct 2026 02:53AM 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 02:42AM UTC coverage: 85.668% (+0.006%) from 85.662%
40807980347

push

github

web-flow
e2e: cover the documented model and tokenizer auto-detection (#955)

Fixes #947

Old behavior: nothing in the e2e tier ran the CLI with
`server.model_name` and `tokenizer.pretrained_model_name_or_path` both
unset — every config pins both, and the unit test for the report
tokenizer calls `sync_tokenizer_config` directly. Deleting that call in
`main.py` kept every test green, and auto-detected runs went back to
reporting with no tokenizer, silently skipping the streamed token-timing
correction.

New behavior: two cases drive the CLI with both fields unset against the
absorber and read the resolved tokenizer path back out of the report
bundle's `config.yaml`. One serves the extracted `google_gemma-3-270m`
directory as the only `/v1/models` id; the other serves the alias
`served-alias` with that directory as `root`, which is what vLLM answers
behind `--served-model-name`. The absorber gained a `root` field for the
second case, defaulting to `model` so the existing parity cases see the
same reply as before.

Changes:
- `e2e/tests/test_model_auto_detection.py`: the two cases plus the
shared config helper.
- `e2e/utils/absorber.py`: `AbsorberServer.root`, reported as
`/v1/models` `root`.

Verification (Windows, local shims for the harness's
`preexec_fn`/`os.setpgrp`, no llm-d-inference-sim needed):
- Both cases pass as committed: `2 passed in 159.84s`.
- With the `sync_tokenizer_config` call in `main.py` removed, both fail:
`AssertionError: bundle config.yaml records no tokenizer path:
tokenizer={}`.
- `ruff format --check` and `ruff check` clean on both files.
- Removing the absorber's `root` makes the alias case fail to start:
`Unable to load a tokenizer for any candidate model name
['served-alias']`.

The `e2e/tests/parity/test_offered_load_parity.py` inference-perf cases
time out here at the harness's 300s limit (request rate reaches 39.96
against a configured 40, then report generation and exit exceed it).
That reproduces with this change stashed, so it... (continued)

13216 of 15427 relevant lines covered (85.67%)

0.86 hits per line

Jobs
ID Job ID Ran Files Coverage
1 40807980347.1 29 Oct 2026 02:53AM UTC 137
85.67
GitHub Action Run
Source Files on build 40807980347
  • Tree
  • List 137
  • Changed 1
  • Source Changed 0
  • Coverage Changed 1
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40807980347
  • 114e7d90 on github
  • Prev Build on main (#40792758044)
  • Next Build on main (#40896748314)
  • 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