AMD has demonstrated how its SGLang Diffusion framework can serve state-of-the-art diffusion models on its GPUs, achieving significant performance improvements. The framework leverages SGLang’s high-throughput scheduler and optimized compute kernels on the ROCm platform, reducing per-step overhead in evaluated configurations. The results show SGLang Diffusion consistently outperforms Hugging Face Diffusers across four tested models, with speedups ranging from 1.5x to 6.3x. The benchmarks were conducted using a single AMD MI350X GPU with identical model weights and inference steps.
The performance comparison highlights the benefits of fused kernels and the AITER attention backend, which help reduce per-step overhead. Distilled models like Z-Image-Turbo see the most significant gains. On AMD ROCm, the AITER backend also handles FP32 to BF16/FP16 dtype casting for encoder components in the diffusion pipeline, ensuring numerically stable inference without requiring user-side code changes. Additionally, the framework supports text-to-image and image editing tasks, producing high-quality visual outputs.
The source provides detailed instructions for reproducing the benchmarks and sample outputs using AMD MI350X GPUs. The framework is designed to support both single-shot generation and benchmarking, with specific commands for environment setup, server launching, and benchmark client execution. These results underscore AMD’s commitment to advancing AI workloads on its hardware.
Source: amd
