sglang-diffusion-benchmark-profile

To measure diffusion denoise latency and collect perf dumps, run the included bench_diffusion_denoise.py script to execute end-to-end benchmarks across multiple models, capturing denoise latency, peak memory, and Torch profiler traces.

Profiling native SGLang diffusion backend performance involves applying end-to-end benchmarks across presets and GPU configurations, using Torch profiler traces and perf dumps to capture performance signals and rank hotspots for kernel-optimization handoffs.

Running preflight checks before benchmarking diffusion models requires enforcing native backend gate validation and environment verification to ensure the native SGLang diffusion backend is active and properly configured before collecting latency metrics.

Yes, you can benchmark end-to-end latency across multiple diffusion models, presets, and GPU configurations to capture denoise latency, peak memory, and performance traces, validating the native diffusion backend performance signals.

Generating Torch profiler traces to identify diffusion kernel bottlenecks is handled by collecting perf dumps during benchmark execution, ranking hotspots from the profiler outputs, and guiding kernel-optimization handoffs for the native backend.

Your diffusion benchmark requires the native SGLang diffusion backend because the Skill enforces preflight checks and native backend gate validation to ensure accurate measurement of denoise latency and reliable collection of profiler outputs.