AMD Quark adds native W4A8 mixed-precision for Wan2.2 video generation on MI350

AMD has published a technical deep-dive on mixed-precision quantization for the Wan2.2 text-to-video generation model using its Quark toolkit on Instinct MI350 GPUs. The work demonstrates a calibration-driven approach that evaluates each transformer block's sensitivity to MXFP4 activation quantization and selectively promotes only the most sensitive blocks from W4A4 to W4A8 precision, rather than applying uniform quantization across the entire 14-billion-parameter model. The dual-expert architecture of Wan2.2-T2V-A14B, which splits denoising between high-noise and low-noise transformer paths, requires per-expert block selection to avoid overfitting quantization policy to one path while undersampling the other.

The study establishes that full W4A4 MXFP4 quantization achieves 0.8520 VBench overall score compared to BF16 reference, with visible quality degradation despite the seemingly close numeric score. Full W4A8 reaches 0.8658, nearly matching BF16, while the mixed-precision policy hits 0.8648 by promoting only sensitive blocks to W4A8. Native W4A8 execution becomes available in AMD Quark 0.13; prior W4A8 results in the study were obtained through quantize-dequantize emulation for quality analysis only.

Quark quantization pipeline and mixed-precision selection

• MXFP4 formats and precision levels: The study uses OCP MXFP4 for weights in both W4A4 and W4A8 configurations, with dynamic MXFP4 activations for W4A4 and FP8 E4M3 static activations for W4A8, using BF16 as the quality ceiling reference.
• Block-level SQNR calibration: The selection pass computes whole-block output SQNR against BF16 for each transformer block under two conditions (plain MXFP4 and MXFP4 with fixed Hadamard rotation plus SmoothQuant), then applies tiered thresholds to decide between plain W4A4, W4A8 promotion, transform-enabled W4A4, or BF16 fallback.
• Dual-expert architecture consideration: Wan2.2-T2V-A14B uses separate transformer and transformer_2 experts for high-noise and low-noise portions of denoising with a boundary_ratio of 0.875, and experiments showed the two experts selected different blocks for promotion, validating per-expert policy decisions.
• Native execution paths: W4A4 uses the Quark/ATER ASM GEMM kernel path on MI350 (gfx950), while W4A8 native execution is introduced in Quark 0.13; all W4A8 quality numbers in this publication used QDQ emulation on Quark 0.12 and are explicitly not performance measurements.
• Fixed rotation selector: The block sensitivity selector uses deterministic fixed Hadamard rotations (R1 for residual streams, R2 for attention head dimensions) followed by SmoothQuant, without learned parameters or training loops, making it faster than SVDQuant-based alternatives while leaving additional quality headroom unexplored.

Why it matters

This work provides a concrete, reproducible methodology for quantizing large diffusion transformers without accepting the quality loss of uniform low-precision formats, which matters for deploying video generation at scale on AMD Instinct hardware. For the ROCm ecosystem, native W4A8 support in Quark 0.13 fills a precision gap between efficient MXFP4 and full BF16, giving developers a tunable quality-efficiency tradeoff that was previously only evaluable through emulation. The per-block, per-expert calibration approach is transferable to other large diffusion models and could influence how quantization tooling handles increasingly common mixture-of-experts and multi-stage architectures in generative AI.