mydubber-org/voxcpm2-int4

VoxCPM2, int4 LMs, prebuilt for myDubber

VoxCPM2 as the myDubber local backend and server workers hold it in memory: built once here so the app downloads and loads it, transforming nothing (no quantization at start, no weight-norm folding, no torchao at runtime).

Built from openbmb/VoxCPM2 at commit 32279effe8c19989596f05d353d1447f51d9e915 by model_builds/voxcpm2.py (dubber). This is a modified copy of VoxCPM2's weights (Apache-2.0 §4(b)):

• the base and residual language models' linears (252 modules) are int4 weight-only, groups of 128 (torchao 0.18.0, Int4WeightOnlyConfig(group_size=128, int4_packing_format="tile_packed_to_4d"), the tinygemm algorithm), stored as plain tensors: <module>.qdata (int32, the layout of torch.ops.aten._convert_weight_to_int4pack on CUDA, inner_k_tiles 8, torch 2.11.0+cu128) and <module>.scale_and_zero (bf16, [in/128, out, 2]). They compute with torch.ops.aten._weight_int4pack_mm on an NVIDIA GPU of compute capability 8.0 or newer; the layout is torch's own, so other torch versions may pack differently (ROCm does);
• the codec's (audio_vae.*) weight norm is folded into plain weights (torch.nn.utils.remove_weight_norm), float32 as upstream runs it;
• everything else is the original's tensors in bf16; the text embedding table, the feature encoder and the DiT are unchanged.

config.json is upstream's plus a prebuilt section listing the int4 modules and the versions; the tokenizer files are upstream's. The audio codec that was audiovae.pth is inside model.safetensors. Loaded by dub_pipeline.fast_voxcpm2.load_prebuilt; the takes are byte-identical to quantizing at load (dubber's S12 harness).

model.safetensors: 2.29 GB, sha256 6bc480d70d58007fdc4b24406ca18b16a552a99fd7fc4daa34be9526c86f705a

Licence

Apache-2.0, as the original (copy in LICENSE). Copyright of the weights: the VoxCPM2 authors (OpenBMB). The modifications above are the only changes; no training or fine-tuning.