sglang-diffusion-benchmark-profile

To benchmark diffusion denoise latency in SGLang, run the bench_diffusion_denoise.py script with your target model to execute an end-to-end run and generate a perf dump for analysis.

You collect torch.profiler traces for diffusion hotspots by running the denoise benchmark, which provides performance traces, hotspot ranking, and maps hotspots back to known fast paths and fusion families.

Yes, you need to export HF_TOKEN as an environment variable when running the denoise benchmark script against gated diffusion models to ensure successful authentication and model access.

The best way to identify performance bottlenecks is using the benchmarking and profiling workflow to generate torch.profiler traces, rank hotspots, and map confirmed kernel work to specialized optimization skills.

Yes, you can compare performance across changes by running the structured denoise latency benchmark and collecting perf data, which enables reliable measurement and comparison across backend modifications.

After profiling identifies diffusion kernel hotspots, you should hand the confirmed kernel work to specialized optimization skills to accelerate performance improvements in the SGLang backend.