metadata
license: apache-2.0
tags:
- comfyui
- diffusion_single_file
Original repo:
https://huggingface.co/lightx2v/Wan2.2-NVFP4-Sparse
LightX2V's NVFP4 → Comfy NVFP4, same datatype, different conventions:
- Nibble swap on U8 weights — Comfy's store_fp4x2 reverses NVIDIA-spec packing
- weight_scale correct as it is
- weight_scale_2 = alpha × input_global_scale — LightX2V stores a CUTLASS post-matmul rescaler, Comfy uses the divisive per-tensor scale
- input_scale = 1 / input_global_scale — same convention flip