[LauraGPT]

Summary

• register the third-party OpenMOSS MOSS-Transcribe-Diarize model with FunASR AutoModel while preserving upstream naming, Apache-2.0 attribution, and immutable revision guidance
• normalize end-to-end ASR + diarization output into FunASR text, raw_text, timestamp, and sentence_info/spk fields without external VAD
• support both local Transformers inference and an existing OpenAI-compatible vLLM transcription service; vLLM client mode bypasses local weight download

Boundaries

• MOSS remains an OpenMOSS model, not a FunASR-owned model
• vad_model and spk_model are rejected because external segmentation would break global speaker identity
• malformed tagged generations remain visible as raw text; the adapter does not invent timestamps or speakers

Verification

• 12 targeted adapter/docs tests passed in the repository environment
• 6 adapter tests passed in the pinned vLLM environment
• 5 unaffected AutoModel tests passed; the excluded merge_thr test fails identically on unmodified main because ModelScope/import dependencies are absent in the host test environment
• exact model revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8 passed a real H100 Transformers smoke test on a 15.1685-second two-speaker probe: 2 segments, S01/S02
• exact vLLM revision 68b4a1d582818e67adc903bf1b8fc5a5447da2fa passed the same probe through AutoModel backend=vllm; the temporary worker was stopped and port 8898 was verified free
• black --check, py_compile, and git diff --check passed

Signed-off-by: LauraGPT 18321252+LauraGPT@users.noreply.github.com