Skip to content

repetition loop - #166

Merged
cjpais merged 5 commits into
mainfrom
fix/decode-degeneration
Sep 26, 2026
Merged

cjpais merged 5 commits into
mainfrom
fix/decode-degeneration

Conversation

@cjpais

@cjpais cjpais commented Sep 21, 2026 •

Copy link
Copy Markdown
Contributor

basically detect degenerate transcriptions and stop inference early and return an error

@cjpais
cjpais merged commit db09681 into main Sep 26, 2026
25 checks passed
cunba-ai pushed a commit to cunba-ai/transcribe.cpp that referenced this pull request Sep 27, 2026
Upstream sync: ggml 0.25.1->0.25.3, release 0.2.4 content, repetition
loop guard (handy-computer#166, new TRANSCRIBE_ERR_OUTPUT_REPETITION=19), better error
code returns (handy-computer#171), koffi async stack size, qwen3_asr lang-hint docs.

Conflict resolutions:
- Version line kept at fork 0.3.0 (superset ABI); upstream's 0.2.4
  version-only bumps dropped. All binding/version spots re-verified with
  scripts/release/prepare.py gates (version sync, FFI drift, Swift pin,
  lock freshness) - all green.
- src/arch/sortformer/model.cpp: kept fork's run_diar_streaming_window
  refactor; applied upstream handy-computer#171 error-code reclassification
  (ggml_init/alloc -> OOM, sched/graph_compute -> BACKEND) at the
  corresponding sites inside the helper.
- src/transcribe.cpp: kept both includes (fork build-info + upstream
  batch-util).
- ABI change is real (new error code 19): abihash moved to
  8c9917fa9596bf0d; Python/TS/Rust bindings regenerated, Swift
  pinnedHeaderHash updated.
- bindings/python/uv.lock + typescript/package-lock.json re-synced.
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.

1 participant