diffusion-optimal-perf

What problem does it solve?

This Skill helps users achieve the best possible performance when running diffusion models with SGLang, addressing issues like slow inference, high VRAM usage, and latency.

Core Features & Use Cases

• Lossless Optimizations: Techniques like torch.compile, warmup, and multi-GPU parallelism that don't sacrifice output quality.
• Lossy Optimizations: Methods such as approximate attention and quantized models that trade a small amount of quality for significant speed or VRAM gains.
• Use Case: A user wants to generate images faster and reduce the memory footprint of a large diffusion model. They can use this Skill's guide to apply a combination of torch.compile, CPU offloading for specific components, and potentially a lossy attention backend to meet their performance targets.

Quick Start

Use the diffusion-optimal-perf skill to apply lossless optimizations for faster diffusion model inference.