[LauraGPT]

Final Publication Acceptance

GitHub Release and PyPI 1.4.19 are both published. Both distributions were downloaded from official PyPI and match the accepted GitHub assets by SHA-256 and size:

• funasr-1.4.19-py3-none-any.whl: 1014074 bytes; SHA-256 31760d1d9bee645f0257e80c00f6872f50d8e398cb41d7a827356121a6bd40f9.
• funasr-1.4.19.tar.gz: 1049723 bytes; SHA-256 87140542bc1e2e8bd5194ce71a9bff9fe9ae5152188adfa2e52c1fd00fdbd75f.

FunASR 1.4.19

Source: 34736414a4a5fd937f2f10563816a8737256115b / tag v1.4.19.

Fixes

• Preserve empty-token Paraformer batch results and timestamp alignment (fix(paraformer): keep one keyed result per utterance when no tokens are predicted #3797/ci: exercise Paraformer empty-token batch regressions #3807).
• Respect an explicit speaker cap in automatic clustering (fix(auto): integrate configurable automatic speaker cap #3809, integrating fix(diarization): honor configurable speaker caps for long recordings #3794).
• Accept and validate JSON-object --hf_overrides in the realtime server (fix(realtime): validate and forward vLLM HF overrides #3810, integrating feat: support hf-overrides for realtime server #3768).

The separate ONNX source fixes are in this Git tag but this release does not publish a new funasr_onnx package. Reporter-specific original-recording confirmation is not claimed. The historical 1.4.18 hotword fix concerns cross-script keyword boundaries, not local-file loading.

Verification Scope

• 141 disjoint release/docs/package/runtime-contract cases passed; 2 local browser checks skipped. Exact new-head and merged-main hosted checks, including browser/build validation, passed separately with actual logs/artifacts accepted.
• The installed candidate wheel passed 56 empty-token/speaker regressions and 13 JSON parser cases. Controlled fixtures are not pretrained ASR quality. All 439 package source files and 13 entrypoints match the accepted source; the newly downloaded final-tag wheel/sdist passed strict Twine, offline installation, actual import and CLI checks.
• Prior completed native GPU evidence was reused by hashing all 314 inventory entries and binding 12 relevant installed modules: synthetic SANM/CIF FP32/BF16 pipelines, random CAM++ CUDA embeddings with native CPU clustering, and native vLLM FP32/BF16 override fixtures. This was not a new GPU run or a pretrained/original-recording quality evaluation.
• An attempted bare offline repository suite retained 5 collection errors and 2 skips: colliding example smoke_test module names, three runtime utility modules parsing required CLI arguments at import, and missing separate funasr_onnx. The full suite is not claimed green; no new exclusions or skipped tests were added to hide these results.
• All 12 payloads passed SHA-256/size and archive-integrity checks. The 2 final-tag Python packages were newly downloaded; the sdist was completed by a verified HTTP Range resume after a transport timeout. Its complete 439-file package-source set, embedded name/version, archive paths and required build/license files were also checked. The 10 unchanged runtime archives reused complete cache bytes freshly hashed against the new asset metadata and runtime-llamacpp-v0.2.6. Runtime binaries were not rebuilt or executed. SHA256SUMS covers all 12 payloads.
• This is scoped release verification, not the deferred Nano nine-dataset benchmark or adoption/growth evidence.

GitHub Release and official PyPI downloads are both accepted; the plan below is retained history.

Historical Pre-Release Plan

Summary

Package three normally merged Python fixes that are absent from 1.4.18:

• Empty-token Paraformer batch results (fix(paraformer): keep one keyed result per utterance when no tokens are predicted #3797, with durable CI selection in ci: exercise Paraformer empty-token batch regressions #3807), retaining MohammadHijjawi97's contribution.
• Automatic speaker clustering respects max_speakers (fix(auto): integrate configurable automatic speaker cap #3809, integrating fix(diarization): honor configurable speaker caps for long recordings #3794 with phoenixray2000 credited).
• Realtime WebSocket --hf_overrides accepts JSON objects and rejects malformed/non-object input (fix(realtime): validate and forward vLLM HF overrides #3810, integrating feat: support hf-overrides for realtime server #3768 and crediting actual author Li Junhui).

Only eight version/test/documentation paths change here. The four README news sections retain three items and the corrected 1.4.18 historical hotword description. Independent ONNX Python package versions are not changed; the repository's ONNX fixes do not imply a new funasr_onnx PyPI release.

Validation

• Release contract RED: 4 failures and 4 passing historical controls before version/docs changes.
• Release/docs/package/runtime contracts: 141 passed and 2 local browser skips, without repeating the first 41 passing cases. New-head hosted build/browser CI is required before normal merge.
• Candidate wheel/sdist: 439 packaged source files byte-equal the candidate, 13 entrypoints, strict Twine, offline install, actual installed import and CLI passed.
• Actual installed modules: 56 empty-batch/speaker regression cases passed (2 plugin-rewrite warnings), plus 13 real JSON-parser cases. Existing controlled encoder/clustering fixtures are not pretrained ASR accuracy tests.
• Prior native L20Z evidence is explicitly reused, not rerun: 32 synthetic SANM/CIF FP32/BF16 cases; synthetic CAM++ CUDA embeddings plus native CPU clustering; native vLLM FP32/BF16 architecture overrides. All 314 supporting inventory entries were rehashed and relevant installed source bytes match the GPU-tested code. No new GPU, model download or Nano experiment is claimed.
• Bounded offline bare pytest stopped during collection with 5 errors and 2 skips: duplicate smoke_test module names across MCP/OpenAI examples; import-time argument parsing in test_cer.py, test_rtf.py, test_rtf_gpu.py; missing separately packaged funasr_onnx in this environment. These failures are retained, not hidden or described as a green full suite.

Exact-head/main/tag CI and final asset acceptance remain separate gates. Existing ten-platform llama.cpp runtime pointer is unchanged. Final release requires the two newly downloaded final-tag Python packages, all twelve payload digests, a verified SHA256SUMS, and PyPI last. Cache reuse for unchanged runtime archives will be explicitly disclosed and matched against new asset metadata and full hashes.

No protection changes, admin bypass, fabricated review or unrelated issue closure. This PR is not itself a completed GitHub/PyPI publication or an adoption-growth claim. Nano work remains deferred.