Diffusion models such as Black Forest Labs’ FLUX.1-dev [1] deliver stunning image quality but demand significant compute and memory bandwidth at inference time. To reduce inference cost without sacrificing image quality, precision-aware quantization techniques have become a critical optimization strategy.
In this blog, we demonstrate how AMD Quark [2], a high-performance quantization library optimized for AMD Instinct™ MI350 GPUs, enables MXFP4 quantization for Diffusers [5] and xDiT [3] FLUX.1-dev image generation. Starting from BF16 eager and torch.compile baselines, we apply Quark native inference with AITER [4] GEMM kernels and show that MXFP4 ASM with torch.compile reaches 1.92× speedup over BF16 eager and 1.41× speedup over BF16 torch.compile, while preserving CLIP quality.