-
Mirror worstchan/WavTTS inference weights as fp16 safetensors
- ema_model_state_dict extracted from model_1200000.pt (10.8 GB
training ckpt), ema_model. prefix stripped, initted/step dropped
- training-only aux_mel_loss.* buffers removed
- floating tensors cast to fp16 (bf16 storage garbles the raw-waveform
ODE state); saved with safetensors save_model
- strict-load verified against MAESTRO's vendored CFM/DiT
- vocab.txt verbatim; config.json points at model.safetensors
Loaded by MAESTRO via backend/ai/models/wavtts.py.
