sglang-diffusion-benchmark-profile

To profile diffusion denoising latency, this Skill automates torch.profiler traces and hotspot ranking to identify bottlenecks in SGLang's native diffusion backend. It maps detected performance limitations to fast paths for targeted optimization.

To run a diffusion latency benchmark, you need the native SGLang diffusion backend, a writable repository, HF_TOKEN for gated models, and idle GPUs configured via diffusion_skill_env.py before executing the benchmark presets.

Yes, you can use checked-in denoise presets and standard perf dumps to generate repeatable GPU benchmarking metrics for local diffusion workloads across supported models in SGLang.

After identifying a kernel bottleneck, the Skill provides workflow handoff guidance to delegate the kernel-optimization work to specialized optimization skills to resolve the detected performance hotspots.