VoxCPM2 — prebuilt weight sets for PyTorch

VoxCPM2 prebuilt for the PyTorch engine of Voicebox-on-Linux, cut exactly the way its panel splits the model: the language part (BaseLM) and the acoustic part (ResidualLM, LocEnc, DiT, projections) each in its own precision. AudioVAE (the head) is always F32.

set	BaseLM	acoustic	size
`q8-f16/`	int8 (`Int8Linear`)	F16	3.7 GB
`q8-f32/`	int8	F32	5.4 GB
`f16-f16/`	F16	F16	5.0 GB
`f16-f32/`	F16	F32	6.7 GB

Why prebuilt: the worker loads only these sets — converting the BF16 original on every load would be slow and spike RAM (both copies in memory). The quantized sets are also faster and lighter on VRAM (q8/f16: ~5.1 GB VRAM, ~3.3 GB RAM on an RX 9070 XT).

Each model.pt is a plain dict of tensors (stan = state dict, bufory = rope buffers, dtype). The engine loads it with torch.load(..., weights_only=True) — no code runs from the file. Made on the CPU with pytorch/workers/voxcpm2/make_weights.py from the repo above — run it yourself if you prefer.

Use

Laid out as Voicebox-on-Linux expects: models/voxcpm-pt/prebuilt/<set>/model.pt (./download-models.sh pytorch-all fetches all four). The config and tokenizer still come from the original repo (models/voxcpm-pt/weights/).

License

Apache 2.0, as the original openbmb/VoxCPM2.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for grruwi/VoxCPM2-pytorch-prebuilt

Base model

openbmb/VoxCPM2

Finetuned

(31)

this model