Skip to content

fix(timestamp): normalize sentence output without punctuation - #3904

Merged
LauraGPT merged 2 commits into
mainfrom
codex/timestamp-no-punctuation-20261115
Nov 15, 2026
Merged

LauraGPT merged 2 commits into
mainfrom
codex/timestamp-no-punctuation-20261115

Conversation

@LauraGPT

Copy link
Copy Markdown
Collaborator

Summary

Integrate #3898 with its original contributor commit preserved as a parent. Both sentence helpers now return string text and honor return_raw_text when punctuation is None or empty. Correct the stale AutoModel malformed-punctuation test to this intentional string contract, retaining timestamp assertions. The existing NumPy matrix selects both regression files; no workflow changes.

Verification

  • Private selected NumPy workflow first run: 437 passes, 27 explicit CUDA-unavailable skips, one stale list-text assertion failure. The corrected assertion fails on baseline and passes on the candidate; unchanged production bytes allow an explicitly aggregated 438 passes/27 skips, NOT a fresh single-run 438 or full repository suite. Original failures are retained.
  • Reuse of earlier timestamp CPU control evidence is bound to unchanged relevant source/test hashes: baseline 2 expected failures/9 passes, candidate 23 passes. Existing usable results were not repeated.
  • Real L20Z native CUDA CIF FP32/BF16 -> timestamp generation -> CPU sentence helpers: 32 baseline contract failures, 32 candidate passes, 1812 CUDA trace events in each phase. Covers Chinese/English, None/empty punctuation, raw text on/off and 0/1000ms offsets. Synthetic tensors, no downloaded weights or pretrained ASR-quality claim.
  • Bare repository pytest was attempted separately in an offline isolated environment. It is NOT represented as a full-suite pass; see the retained scope below.
  • Signed+DCO integration, inline code review, no self-approval or protection bypass. New exact-head and merged-main CI are separate gates. Nano remains paused.

Full-suite limitation: Bare repository pytest stopped during collection: 2 unchanged standalone funasr-onnx test modules could not import that separate package in the private environment (6 skips). No test-body/full-model-suite pass is claimed. The selected root-package NumPy workflow and native timestamp CUDA contract are independently validated. The author-reported unrelated AutoModel failure was not reached by this collection attempt.

kushaldave002 and others added 2 commits November 13, 2026 21:59
…nctuation

With no punctuation ids, timestamp_sentence and timestamp_sentence_en
returned the text as a list of tokens and ignored return_raw_text. Let
the regular loop build the single sentence so text is a string and
raw_text is honored.
Preserve the original #3898 author commit and current main with real CUDA and selected suite evidence.

Signed-off-by: LauraGPT <zhifu.gzf@alibaba-inc.com>
@LauraGPT
LauraGPT merged commit af74d4c into main Nov 15, 2026
2 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