VoxCPM2 AudioVAE — safetensors conversion
This repo contains a straight format conversion of openbmb/VoxCPM2's audiovae.pth (PyTorch pickle) to audiovae.safetensors. No weights were changed — this is torch.load(audiovae.pth, weights_only=True) followed by safetensors.torch.save_file(), nothing else.
Why
Crane (a Rust/candle inference engine) loads VoxCPM2's AudioVAE via candle's safetensors loader, which has no PyTorch pickle reader. Rather than requiring every Crane user to run the conversion themselves, this repo hosts the converted file directly.
Usage
Download audiovae.safetensors and place it alongside the rest of the openbmb/VoxCPM2 checkpoint files (model.safetensors, config.json, tokenizer.json, etc.) in the same directory before loading with Crane.
License
Apache-2.0, inherited from the original openbmb/VoxCPM2 checkpoint. All credit for the model itself goes to OpenBMB — this repo only hosts a converted copy of one file for convenience.