MOSS-TTS-Nano-100M — GGUF (Q8_0) for RapidSpeech.cpp / ggml-CUDA

Q8_0 GGUF conversion of OpenMOSS-Team/MOSS-TTS-Nano-100M for on-device inference on Jetson Nano gen1 (Maxwell sm_53) via ggml-CUDA.

• moss_nano_full.gguf — merged AR model + codec decoder, Q8_0 (~138 MB from 440 MB fp32).
  • AR: GPT-2 12-layer global + 1-layer local decoder + 16 audio codebook heads (interleaved RoPE, gelu_new).
  • Codec: MOSS-Audio-Tokenizer-Nano decoder + 16-way RVQ (weight_norm reconstructed) → 48 kHz.
• moss_nano.gguf — AR model only. moss_codec.gguf — codec only.

Verified (torch-free, vs the deployed ONNX): global transformer prefill+decode MSE 1.4e-6, local decoder 14/16 argmax (Q8), codec round-trip corr 0.95–0.998. Performance (Jetson Nano GPU, ggml-CUDA): Q8 RTF 0.86 (full 12L) / 0.50 (4L student); ~0.35 with the custom sm_53 matvec kernel. RTF floor ~0.12.

Runtime: vieenrose/RapidSpeech.cpp (branch jetson-nano-gen1), arch moss_tts_nano. Converters in scripts/convert_moss_*_to_gguf.py.