What changed
AMD has introduced mixed-precision quantization for the Wan2.2 video generation model, leveraging its Quark framework to optimize performance and quality on AMD Instinct GPUs. The approach involves using MXFP4-weight quantization with mixed activation precision, promoting sensitive blocks from MXFP4 W4A4 to W4A8. This calibration-driven method aims to balance efficiency and quality, achieving a VBench overall score of 0.8648 for Wan2.2-T2V-A14B.
What this means in practice
The mixed-precision quantization allows for efficient execution while preserving video generation quality. For the Wan2.2-T2V-A14B model, Full W4A4 MXFP4 is efficient but shows a visible quality drop, whereas Full W4A8 reaches the highest measured quantized quality. Native W4A8 execution is available in AMD Quark 0.13, and the calibration process identifies which blocks benefit most from higher precision. This approach is tailored for the dual-expert structure of the model, where different blocks require varying levels of precision to maintain performance.
Limitations and availability
The W4A8 results are based on quantize-dequantize (QDQ) emulation and are reported as a quality study, not a performance result. Native W4A8 execution is available starting with AMD Quark 0.13. The calibration process uses specific transforms like fixed rotation plus SmoothQuant to identify sensitive blocks, and the results are measured using VBench as the quality metric. The experimental setup uses a single AMD Instinct MI350 GPU and specific versions of Docker and Quark for consistency.
