thepatch/bs-roformer-viperx-317-GGUF

BS-RoFormer (viperx ep_317) GGUF

No license is stated for these weights anywhere upstream. viperx's checkpoint is distributed through Ultimate Vocal Remover's model repository without one. It is republished here in GGUF form with that caveat. Do not assume any license, and if you hold the rights and want it removed, open a discussion on this repository.

viperx's Band-Split RoFormer vocal model (model_bs_roformer_ep_317_sdr_12.9755), converted to GGUF for stems.cpp, a C++/ggml stem separator that runs on CPU, CUDA, Vulkan and Metal with no Python. It estimates vocals; the instrumental is the mix minus the vocals.

file	size
bs_roformer_viperx_317-0.2B-v1.0-F32.gguf	609 MiB

Download and run

git clone --recursive https://github.com/betweentwomidnights/stems.cpp && cd stems.cpp
./build.sh vulkan                          # or cpu | cuda | metal   (Windows: build.cmd)
./models.sh bs_roformer_viperx_317
build-vulkan/bin/stems-split -m models/bs_roformer_viperx_317-0.2B-v1.0-F32.gguf -i song.wav -o stems/

stems-server finds it by model name ("model": "bs_roformer_viperx_317").

Parity with PyTorch

Each figure is the vocal-stem SNR of the C++ output against ZFTurbo's MSST BSRoformer run in float32 on CPU, for a 20 s clip (demucs' test.mp3), as seg (the first 8 s chunk) / full (chunked demix_track). Every result passes stems.cpp's parity check. Measured 2026-10-02/03 on stems.cpp with ggml f30f0cdc. CPU, CUDA and Vulkan: Core Ultra 9 275HX / RTX 5070 Laptop. Metal: Apple M4, against torch refs dumped on the M4; that machine's CPU gets 61.8 / 50.6 against the same refs, so Metal is 1.8 dB under it on the full track, which is float32 accumulation over 12 layers.

encoding	CPU	CUDA	Vulkan	Metal	20 s clip (Vulkan, RTX 5070)
F32	69.6 / 65.2	69.5 / 62.4	69.6 / 63.1	61.8 / 48.8	14.6 s

F16 is not published for this model. It was measured (44.2 / 34.9 dB on CPU, 45.1 / 36.3 on CUDA, 44.0 / 31.9 on Vulkan). That clears stems.cpp's cosine check, but it loses about 30 dB and lands where Vulkan was before ggml learned to honour GGML_PREC_F32, a level stems.cpp treated as a bug. Mel-Band (Kim) and HTDemucs lose much less at F16 and are published at both encodings.

Sources and license

• Upstream checkpoint: TRvlvr/model_repo release all_public_uvr_models, file model_bs_roformer_ep_317_sdr_12.9755.ckpt, sha256 5b84f37e8d444c8cb30c79d77f613a41c05868ff9c9ac6c7049c00aefae115aa. Trained by viperx.
• Configuration: ZFTurbo's Music-Source-Separation-Training (MIT), configs/viperx/model_bs_roformer_ep_317_sdr_12.9755.yaml.
• Converted with stems.cpp's tools/convert_roformer.py --preset viperx. Tensors are renamed, the band layout is written as plain index lists, and the rotary frequencies are checked and stored as a theta. The architecture and weights are otherwise unchanged.
• No license. There is no LICENSE file here because upstream provides none; NOTICE records the provenance and the caveat. SHA256SUMS lists the GGUF.