Repository navigation
Release 0.50.22
Install
Docker
docker pull ghcr.io/jedwards1230/earmark:0.50.22Helm
helm install earmark oci://ghcr.io/jedwards1230/charts/earmark --version 0.50.22v0.50.22 Release Notes
This patch release changes the ASR pilot runner (runner/pilot.py) to load models offline from local Hugging Face snapshots, and it lowers the default Granite window length. It also adds step-run tracking for pilot runs and a few new CLI options.
Breaking Changes
- Granite default window reduced from 240 s to 90 s, and the hard maximum from 540 s to 120 s. Per the runner docs, longer windows caused re-reading and dropped speech. Canary is unchanged (38 s default, 40 s max).
- Migration: if you set
PILOT_WINDOW_SECfor Granite to a value above 120, the run now fails with a config error. Set it to 120 or lower. --bench-max-secno longer defaults to 1800. It now defaults tomax(--bench-batches) × 2 × longest window. Pass it explicitly to keep a fixed cap.
- Migration: if you set
- Offline model loading requires prefetched snapshots. Runs set
HF_HUB_OFFLINE=1and read model directories from the local HF cache. They no longer download on demand. A missing snapshot or file fails the load with an error naming theasr_pilot_windows_<role>_allow_patternsvariable to fix.- Migration: prefetch the snapshots before running. Run
python pilot.py --backend <granite|canary> --print-prefetch-specto list the repos, revisions, and allow-patterns that pilot.py needs. - Canary also needs the
Qwen/Qwen3-1.7Bsnapshot (revision70d244cc86ccca08cf5af4e1e306ecf908b1ad5e, or override withPILOT_QWEN_REVISION).
- Migration: prefetch the snapshots before running. Run
--benchbehavior changed. It gathers windows across the--asintracks instead of using only the first track, and it ignores--limit-tracks.
Features
- Offline model loading. Models load from local snapshot directories in the Hugging Face cache, and
--print-prefetch-specprints the required snapshots as JSON. Canary's LLM config is pointed at the local Qwen snapshot, so no repo lookup happens at load time. - Pilot runs appear in run tracking. A real run writes one
step_runsrow (stepasr_pilot), so it shows in the dashboard's Running now panel and inGET /api/v1/runs. It reports progress, and it is closed asdone,failed, orcancelled. Dry runs,--bench, and--self-checkrecord nothing. --bench-windowssweeps several window lengths (for example60,90,120) in one bench run. Bench output now includes per-row degenerate rate and word completeness.--no-repeat-ngram-size N(Granite, default off) is available for experiments.
Improvements
- Granite stops each batch row on its own once it repeats an 8-token n-gram three times, so one re-reading row no longer holds the whole batch.
- Degenerate-window detection now also flags re-reads, meaning the same 8-word run three times or the same 16-word run twice. It also flags windows with more than 4 words per audio second as
overlong. - Special tokens that leak into the output (such as
<|fim_middle|>) are stripped from the text of both backends.
Commits
- chore(earmark): bump chart version to 0.50.22 (f79e1a6)
- fix(runner): load pilot models from local HF snapshots, offline (Canary self-check) (#223) (342f6ab)
Full Changelog: v0.50.21...v0.50.22