Faster Diffusion on Blackwell: MXFP8 and NVFP4 with Diffusers and TorchAO – PyTorch

MXFP8 and NVFP4 microscaling quantization formats natively supported by NVIDIA's Blackwell architecture (B200 GPUs) deliver significant inference speedups for diffusion models. Using diffusers and TorchAO, benchmarks on Flux.1-Dev, QwenImage, and LTX-2 show up to 1.26x speedup with MXFP8 and up to 1.68x with NVFP4 compared to BF16 baselines, while also reducing peak memory by up to ~3.5x. The post covers selective quantization strategies (skipping small or accuracy-critical layers), CUDA Graphs via torch.compile reduce-overhead mode to cut CPU overhead at small batch sizes, and LPIPS as a perceptual accuracy metric. Code and reproducible recipes are provided for all three models.