sglang-diffusion-performance

To find the fastest diffusion configuration for your GPU, benchmark native SGLang and Diffusers backends. Comparing parallelism schemes, attention backends, and caching decisions identifies the most VRAM-efficient setup for your specific model and hardware.

Yes, you can benchmark SGLang vs Diffusers backends directly to compare performance. The skill analyzes native SGLang pipelines against Diffusers backends across single- and multi-GPU scenarios to determine the fastest generation flags.

You need a CUDA-capable environment, access to SGLang generate with either --backend sglang or --backend diffusers, and the ability to produce performance dumps. This allows the skill to analyze and identify optimal VRAM configurations.

Optimize diffusion VRAM usage across multiple GPUs by comparing different parallelism schemes. The skill benchmarks multi-GPU configurations against attention backends and caching decisions to find the most efficient setup within your VRAM budget.

Yes, SGLang supports different attention backends for diffusion models. You can benchmark these attention backends alongside caching decisions and parallelism schemes to identify the fastest and most VRAM-efficient configuration for your setup.