-
Mirror OpenMOSS-Team/MOSS-TTS-Nano-100M @ 44502f80dbf9 + OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano @ 6aa02b01e445 for MAESTRO (nano/)
- model/: upstream pytorch_model.bin → model.safetensors (tensors verbatim bf16, the 17 storage-shared tied heads dropped and re-tied at load); provenance in model/sha256.json
- codec/: re-sharded by top-level module (encoder / decoder / quantizer); provenance in codec/sha256.json
- HF remote-code .py files dropped: MAESTRO vendors the graphs (ai/moss_tts_vendor), never trust_remote_code
Loaded by MAESTRO via backend/ai/models/moss_tts_nano.py (CPU lane, whole codec resident).
-
Mirror OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 @ be7766a6735b + OpenMOSS-Team/MOSS-Audio-Tokenizer-v2 @ f6e20e543b33 for MAESTRO (local_v15/)
- model/: safetensors + tokenizer/config files VERBATIM (sha256 vs upstream LFS)
- codec/: re-sharded by top-level module (encoder / decoder / quantizer); provenance in codec/sha256.json
- HF remote-code .py files dropped: MAESTRO vendors the GitHub graphs (ai/moss_tts_vendor), never trust_remote_code
Loaded by MAESTRO via backend/ai/models/moss_tts.py (codec halves sequenced on/off the GPU). -
initial commit
