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

kubernetes-sigs / inference-perf / 40729774242
87%

Build:
DEFAULT BRANCH: main
Ran 28 Oct 2026 05:01PM 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

28 Oct 2026 04:49PM UTC coverage: 85.539% (+0.3%) from 85.239%
40729774242

push

github

web-flow
Bound ShareGPT memory use across repeated dataset passes (#934)

Fixes #923

ShareGPT's top-level JSON array was materialized despite
`streaming=True`, and `itertools.cycle` retained decoded rows. The
loader now parses incrementally and reopens the source for each pass.
Parsing turns as objects also fixes Chat and Anthropic Messages on the
default file; `datasets` had returned each turn as a JSON string.

Changes:
- Download the original Hub file through `hf_hub_download` within
`data.load_timeout`, then use the same incremental disk reader as local
input. Repeat passes reuse the cached file without an HTTP response
staying open between stages.
- Parse JSON arrays and JSON Lines with `ijson`; preserve record order,
startup priming, filtering, and lazy worker loading. Close files between
passes and reject empty/malformed input.
- Cover bounded memory and reads, two-pass order, cache-only iteration,
download errors, startup/reload timeouts, and all three APIs. Bound read
assertions by `max(65536, io.DEFAULT_BUFFER_SIZE)` for Python 3.14.
- Document the behavior and drop the `sharegpt-hub` xdist group. The
hook repair and e2e annotations are in #936 and are excluded from this
PR.

Regression evidence:
```
Original loader memory regression:
E assert 52385889 < ((2 * 1024) * 1024)

Previous PR head on Python 3.14:
E assert 131072 <= 65536

Previous remote-stream reader after a simulated disconnect:
OSError: connection lost between passes
```

With the original 672,837,942-byte Hub file (94,145 records), measured
peak RSS was 214.3 MiB through startup and two complete passes, versus
3,679.7 MiB for the baseline in fresh macOS arm64/Python 3.12.15
processes. Every complete pass matched the baseline digest over ordered
IDs and decoded conversations. The [reviewer independently
verified](https://github.com/kubernetes-sigs/inference-perf/pull/934#issuecomment-6448161805)
120,000 matching completion requests across the pass boundary, 6,179 to
234 MiB peak RSS, a... (continued)

19 of 20 new or added lines in 1 file covered. (95.0%)

13155 of 15379 relevant lines covered (85.54%)

0.86 hits per line

Uncovered Changes

Lines Coverage ∆ File
1
88.44
34.3% inference_perf/datagen/dataset/hf_sharegpt_datagen.py
Jobs
ID Job ID Ran Files Coverage
1 40729774242.1 28 Oct 2026 05:01PM UTC 137
85.54
GitHub Action Run
Source Files on build 40729774242
  • Tree
  • List 137
  • Changed 1
  • Source Changed 1
  • Coverage Changed 1
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40729774242
  • 26f70ddc on github
  • Prev Build on main (#40728476156)
  • Next Build on main (#40792758044)
  • 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