Noticed there was no quant available for this checkpoint and thought I'd give it a go. There isn't an original FP16 to work with so did what I could with whats available. Done a few quick tests, seems to work as it should!The Wan 2.2 FP8-scaled Safetensors models were converted to GGUF Q4_K_M format using ComfyUI-GGUF and a patched llama.cpp quantizer.The original FP8 weights were dequantized to FP16, with their associated FP8 scaling factors applied during conversion.The resulting FP16 GGUF was then quantized to Q4_K_M using llama.cpp.Wan's required 5D <code>patch_embedding.weight</code> tensor was restored after quantization so the model retains the correct Wan architecture.The final models were checked to ensure the required tensors and Wan architecture were present.The resulting files are Q4_K_M GGUF versions of the original Wan 2.2 FP8-scaled models. No additional model training or fine-tuning was performed.

Files with SHA256: 2d5e6b967004f07800205f8549903e8a3014f6a662fe20b5f79291a69d00cff7