Skip to content

fix(auto): integrate configurable automatic speaker cap - #3809

Merged
LauraGPT merged 1 commit into
mainfrom
fix/speaker-cap-integrated-20261031
Oct 31, 2026
Merged

LauraGPT merged 1 commit into
mainfrom
fix/speaker-cap-integrated-20261031

Conversation

@LauraGPT

Copy link
Copy Markdown
Collaborator

Summary

Integrates @phoenixray2000's #3794 on current main. All seven reviewed changes are preserved: the top-level max_num_spks argument (default 15), a cap on automatic UMAP/HDBSCAN clustering, known-count precedence, validation, bilingual docs, and regression coverage. The original fork/head is untouched and the signed commit retains contributor credit.

The original head has six successful checks and one failed browser check. Its downloaded browser log contains runner-startup information only, so a specific root cause is not claimed. Current main already pins the browser runner to Ubuntu 24.04. This integration requires its own successful CI, not a rerun or bypass of the old failed head.

Validation

  • New cap tests reproduced 41 expected failures on the pre-fix source. The candidate passed 66 focused CPU tests, no skips, with one sklearn future-default warning.
  • A real L20Z ran a random-initialized, reduced CAMPPlus model in FP32, batch 32, over 2048 unique synthetic feature rows. The first-batch profiler recorded 972 CUDA kernel events. Actual CPU UMAP/HDBSCAN produced eight clusters; the cap reduced them to three, with cap 1/3/5 and the spectral path checked separately.
  • Full hashes of the durable receipts, inputs, model parameters, RNG state, embeddings and labels were verified. These are synthetic execution and clustering-boundary tests, not pretrained diarization accuracy, BF16, the complete repository suite, or Nano evidence. No model weights were downloaded.
  • The only changes from the tested base to current main are three unrelated ONNX files. All seven speaker-cap files and dependencies are byte-identical to the tested candidate, and the existing ONNX result is preserved. Completed experiments were not rerun.
  • New-head and merged-main CI/artifacts must be accepted before completion. No release, PyPI upload, protection change or administrator bypass is included. fix(diarization): honor configurable speaker caps for long recordings #3794 will be closed as superseded only after integration, not described as a merged original commit.

Integrate the reviewed #3794 speaker-cap changes on current main, preserving the original fork and contributor credit.

Co-authored-by: phoenixray2000 <phoenixray2000@gmail.com>

Signed-off-by: LauraGPT <zhifu.gzf@alibaba-inc.com>
@LauraGPT
LauraGPT merged commit a367557 into main Oct 31, 2026
7 checks passed
@LauraGPT

Copy link
Copy Markdown
Collaborator Author

Normal signed+DCO merge completed; the final-main tree matches the verified 4158-blob integrated source. New-head and final-main CI succeeded, including the browser check. Actual artifact/log bytes and the source-commit markers were verified. Each NumPy matrix includes 50 cluster tests and the six earlier empty-batch regressions, with no failures or skips.

Local scope remains 66 focused CPU tests plus synthetic reduced-CAMPPlus CUDA/native-clustering validation, not pretrained diarization quality, BF16 or the full suite. #3794 is closed as superseded with credit and its original fork/head untouched. No administrator bypass, protection change, tag, release or PyPI upload was used.

@LauraGPT LauraGPT mentioned this pull request Nov 1, 2026
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