-
Mirror openbmb/VoxCPM2 for MAESTRO
- model.safetensors bf16 (smoke-gate verdict; upstream
ships bf16 — verbatim copy when unchanged, sha256-checked)
- audiovae.pth converted to audiovae.safetensors via save_model
(weight_norm weight_g/weight_v aliasing — shared-tensor safe)
- tokenizer_config.json rewritten without the list-form
extra_special_tokens key; tokenization_voxcpm2.py dropped
(inference uses plain LlamaTokenizerFast, no trust_remote_code)
- from_local load verified against MAESTRO's vendored voxcpm2
inference graph before upload
Loaded by MAESTRO via backend/ai/models/voxcpm2.py.
