#11995 Gallery model audio-cpp-sortformer-diarization fails to install: SHA mismatch, reproducible across retries

LocalAI version

v4.9.0 (localai/localai:latest-gpu-nvidia-cuda-12, revision f7ad3f70eb5d8a0ddf80e08557f0d7df28cf032e)

Describe the bug

Installing the gallery model audio-cpp-sortformer-diarization always fails with a SHA mismatch. The downloaded file is consistent across attempts, so the checksum recorded in the gallery metadata appears to be wrong rather than the download being corrupted.

error: SHA mismatch for file "/models/audio-cpp/sortformer-diar-4spk-v1-q8_0.gguf"
( calculated: 4fa6a3e30c4a1c6cc1da455268806edc93432ca1e1b5b6923e942be8e6e479ff
 != metadata:  84744b047296daf2177aed9dd2735497152ff83f49ec170d2a6a18cbcb93ef19 )

I retried after deleting the partial and completed file, and the calculated hash was byte-for-byte identical the second time. That points at the metadata rather than a transfer problem.

Because the install fails at the verification step, the model is never usable and /v1/audio/diarization has nothing to run.

To Reproduce

Install the cuda12-audio-cpp backend (this succeeds).
POST /models/apply with {"id": "localai@audio-cpp-sortformer-diarization"}
Poll the job. It ends with the SHA mismatch above.
Delete /models/audio-cpp/sortformer-diar-4spk-v1-q8_0.gguf and repeat: the calculated hash is the same again.

Expected behavior

Either the gallery metadata should carry the hash of the file actually served, or the upstream artifact should be restored to the version the metadata describes.

Additional context

Every other model I installed the same way worked, including granite-embedding-107m-multilingual, whisper-large-turbo-q5_0, insightface-antelopev2, kokoro, voice-detect-campplus, voice-detect-ecapa-tdnn, voice-detect-emotion-wav2vec2 and voice-detect-age-gender-wav2vec2. So this looks specific to the sortformer gallery entry rather than a general problem with checksum verification.