thepatch/musicgen-large-GGUF

musicgen-large GGUF

GGUF conversion of facebook/musicgen-large, Meta's MusicGen large (3.3B parameters), for audiocraft.cpp's native musicgen-server, which runs gary4local's C++ gary service.

Pick one precision of the language model and put it in a folder with the EnCodec and T5 files. If more than one precision is in the folder, the server runs the most precise. Every quantized tensor round-trips against the F16 conversion at cosine 0.9999 or better for Q8_0 and 0.9966 or better for Q4_K_M.

On an RTX 5070 Laptop GPU (8 GB, CUDA), a guided 30-second continuation from a 6-second prompt with a MusicGen large finetune peaked at about 5.7 GiB of VRAM at Q4_K_M and 6.9 GiB at Q8_0, desktop included, and took about 43 seconds with either. F16 spilled past the card and took over 100 seconds, so it is for GPUs with 12 GB or more.

File	Size
musicgen-large-3.3B-v1.0-F16.gguf	6,213 MiB	F16; the reference
musicgen-large-3.3B-v1.0-Q8_0.gguf	3,302 MiB	Q8_0 throughout
musicgen-large-3.3B-v1.0-Q4_K_M.gguf	2,054 MiB	Q4_K, with the FFN down projection, cross-attention K/V, codebook embeddings and output heads at Q6_K
encodec-32khz-v1.0-F32.gguf	221 MiB	MusicGen's 32 kHz EnCodec
t5-base-encoder-0.1B-v1.0-F16.gguf	210 MiB	T5-base text encoder for descriptions

$env:AC_DEVICE = "cuda"
musicgen-server --port 8000 --models-dir C:\path\to\this\folder

The server lists the model in /api/models as facebook/musicgen-large and accepts that name as model_name, the same routes gary4juce uses with the Python service. See SERVICES.md.

Sources and licensing

• Language model: facebook/musicgen-large revision 15ccdc92099879e47b6da12c350cdb71d4eab3ca (state_dict.bin), converted to F16 and quantized with audiocraft.cpp's ac-quantize.
• EnCodec: MusicGen's 32 kHz compression model, facebook/encodec_32khz.
• T5 encoder: google-t5/t5-base revision a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1, Apache 2.0.

MusicGen's weights are released by Meta under CC-BY-NC 4.0, and this conversion inherits that license: non-commercial use only. All credit for the model goes to Meta's AudioCraft team (Copet et al., 2023). T5-base is Apache 2.0. Each GGUF records its own source and license in its general.* metadata.