sglang-diffusion-performance

What problem does it solve?

Optimizes diffusion pipeline performance for SGLang by selecting the fastest flags and configurations.

Core Features & Use Cases

• Guided optimization: Recommends lossless and lossy optimization options (e.g., Torch compile, offload strategies, attention backends) to improve throughput with minimal quality impact.
• Benchmark-driven tuning: Provides strategies to compare configurations and identify bottlenecks for single-GPU and multi-GPU deployments.
• Use case: When deploying SGLang diffusion models on Wan2.2 or Flux hardware, use this skill to reach the lowest latency within a given VRAM budget.

Quick Start

Configure a diffusion run to maximize throughput on Wan2.2 with 8 GPUs using torch.compile and offload settings while staying within VRAM limits.