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

kubernetes-sigs / inference-perf / 43302222367
87%

Build:
DEFAULT BRANCH: main
Ran 12 Nov 2026 08:51PM UTC
Jobs 1
Files 142
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

12 Nov 2026 08:37PM UTC coverage: 86.464% (+0.05%) from 86.411%
43302222367

push

github

web-flow
Support multimodal output in reports and /metrics (#946)

Fixes #945

Old behavior: responses carried text only. `ResponseMetrics` held output
tokens and the server usage block, so an endpoint that returns media had
nowhere to record it, and /metrics exported no media in either
direction. Media size fields defaulted to 0, so a field the client never
measured would be averaged into the size distributions as 0.

New behavior: `ResponseMetrics` records image, video and audio with the
same classes `RequestMetrics` uses. The summary report adds
`output_image`, `output_video`, `output_audio` and
`throughput.output_{images,videos,audios}_per_sec` when a response
returns media. A text run's per-request JSON is unchanged. Its report
gains one key, `video.seconds`, null. /metrics exports media counts and
sizes for both directions. An unmeasured field is None and stays out of
every distribution. No endpoint returns media yet, so the output side
stays empty until the follow-up endpoint PRs.

Changes:

* Measurement fields on `Image`, `Video` and `Audio` are `Optional`,
default None. `Video` gains `seconds`. No input spec fills it yet, since
none declares a frame rate.
* Remote specs return all-None records. They are not wired into the
materializer yet, so no report changes today.
* `ResponseMetrics` gains `image`, `video` and `audio`, dropped from the
serialized record when None.
* `summarize_requests` builds every media block through one
`summarize_media`, which skips None. Response blocks appear only when
some response returned media. The input `video` block gains `seconds`,
null today.
* The BR adapter skips None the same way, including a video with no
frame count in the pixels-across-frames statistic. Response media stays
out of the BR until llm-d/llm-d-benchmark#2127 ships in a release, since
v0.2.1 rejects unknown keys.
* New `sets/media.py`: `inference_perf_{input,output}_media_total`,
`_media_bytes`, `_media_pixels` and `_media_seconds` by stage and
modal... (continued)

96 of 96 new or added lines in 8 files covered. (100.0%)

1 existing line in 1 file now uncovered.

13810 of 15972 relevant lines covered (86.46%)

0.86 hits per line

Coverage Regressions

Lines Coverage ∆ File
1
79.17
-1.04% inference_perf/datagen/base.py
Jobs
ID Job ID Ran Files Coverage
1 43302222367.1 12 Nov 2026 08:51PM UTC 142
86.46
GitHub Action Run
Source Files on build 43302222367
  • Tree
  • List 142
  • Changed 13
  • Source Changed 12
  • Coverage Changed 9
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #43302222367
  • c361a89f on github
  • Prev Build on main (#43286339232)
  • Next Build on main (#43353282186)
  • 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