sglang-diffusion-benchmark-profile

Benchmarking diffusion denoise latency in SGLang requires invoking the bench_diffusion_denoise.py script. This process leverages the native backend to collect detailed denoise latency, end-to-end latency, and peak memory metrics for a chosen model.

The best way to profile diffusion bottlenecks is using this skill to collect perf dumps and rank hotspots. It maps identified hotspots to known fast paths and validates the baseline denoise latency before handing off confirmed kernel work to specialized optimization skills.

Yes, you must export HF_TOKEN when benchmarking gated models with SGLang. The skill's preflight helper validates environment setup and exposes required tokens to ensure native-backend guarded benchmarking can securely access the gated model repository.

Perf-dump collection works by enabling torch.profiler during the benchmark execution of the SGLang diffusion backend. This captures detailed performance traces, allowing you to surface performance bottlenecks and generate consistent comparison data across multiple runs.

No, you cannot use this profiling skill without the SGLang native backend. The skill explicitly requires the native SGLang diffusion backend to ensure measurements come from the intended execution path and to enable accurate perf-dump collection.

You need to run the preflight helper to locate the SGLang repo root and verify write access before profiling diffusion performance. This ensures the environment is correctly configured for safely generating perf dumps and collecting latency metrics.