Skip to content

fix(auto): preserve zero-duration results and cover them in CI - #3902

Merged
LauraGPT merged 2 commits into
mainfrom
codex/zero-speech-time-ci-20261115
Nov 14, 2026
Merged

LauraGPT merged 2 commits into
mainfrom
codex/zero-speech-time-ci-20261115

Conversation

@LauraGPT

Copy link
Copy Markdown
Collaborator

Summary

Integrate #3896 while preserving its original contributor commit as a parent. Keep native model results when short input reports zero speech time; both per-batch and average RTF handle zero duration. Add the original regression file to pull-request/main filters and the existing NumPy 1.26.4/2.4.0 commands.

Verification

  • Private numpy-wpe: 436 passed; 27 skipped.
  • Private kws: 100 passed; 0 skipped.
  • Private moss-local-http: 117 passed; 0 skipped.
  • The 27 private NumPy skips are explicit CUDA-unavailable cases in the CPU-only suite; not classified as passes. This is the selected workflow scope, not the full repository suite.
  • Separate real L20Z native FsmnVADStreaming/WavFrontendOnline FP32 validation uses identical synthetic weights/audio: baseline six cases includes two expected ZeroDivisionErrors; candidate six cases pass. Both progress-bar modes and 160/480/960-sample boundaries are covered; positive controls execute CUDA kernels (204 trace events per phase). No pretrained recognition-quality claim or model download.
  • Initial harness assumptions (480-sample metadata and AutoModel key shape) were corrected and retained. Initial broader private suite had five missing-pytorch_wpe errors; read-only existing private dependency binding corrected the environment before acceptance, without changing shared packages or production.
  • MOSS acceptance joins 115 original passes with two real localhost HTTP fixture passes on identical source. The initial isolated-network failures and denied private-loopback setup are retained; only these two fixtures were rerun using existing loopback with ephemeral ports and offline model flags. No privilege or host network configuration was changed; this is not a fresh single-run 117-case result.
  • CI coverage checks fail before the workflow edit and pass afterwards. Code/test bytes are unchanged from the private suites; exact new head CI and eventual merged-main CI remain separate acceptance gates.
  • Signed+DCO integration, inline code review only, no self-approval, protection change, forced push or bypass. Nano remains paused.

sohailshk and others added 2 commits November 13, 2026 10:05
Input shorter than one feature frame (about 25 ms) makes the model report
batch_data_time == 0, and the rtf statistic then raised ZeroDivisionError
after the model had already returned its (empty) result.
Preserve the original #3896 author commit and current main; exercise the added regression file in both existing NumPy matrices.

Signed-off-by: LauraGPT <zhifu.gzf@alibaba-inc.com>
@LauraGPT
LauraGPT merged commit c2107cd into main Nov 14, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants