FartingBackwards/wan2.1-vace-14b-int8-convrot

Wan2.1-VACE-14B INT8 ConvRot

Converted from original FP16 weights (Wan-AI/Wan2.1-VACE-14B, via Comfy-Org repack).

Converted with silveroxides/convert_to_quant:

ctq -i <source>.safetensors -o <output>.safetensors \
    --int8 --scaling_mode row --simple \
    --convrot --convrot-group-size 256 \
    --wan --exclude-layers "vace_blocks\.\d+\.(before|after)_proj" \
    --comfy_quant --save-quant-metadata

VACE proj layers kept unquantized

VACE architecture adds 8 context blocks (vace_blocks.0–.7) that inject into the main transformer via before_proj (block 0 only) and after_proj (all 8 blocks). These are zero-initialized and only lightly adapter-tuned on a frozen base, so their weight magnitudes are much smaller than a normally-trained layer. Row-wise INT8 quantization at the same settings as the rest of the model risks burying that signal in quantization noise, silencing VACE conditioning while the base model still generates normally. These 9 tensors are excluded and kept at source precision; the rest of vace_blocks (attention/FFN/norms) is quantized the same as the base model.

License

Apache 2.0, inherited from the base weights (Wan-AI).