sglang-diffusion-benchmark-profile

Profile diffusion denoise latency by running the native SGLang diffusion backend benchmark preset, which captures torch.profiler traces and hotspot rankings to locate kernel bottlenecks.

Identify diffusion hotspot bottlenecks using checked-in perf dumps and native benchmark scripts that map latency hotspots to fast paths or kernel optimization work.

Yes, profiling diffusion denoise latency requires CUDA-enabled GPUs, HF_TOKEN environment setup, and access to preflight utilities to validate native backend paths and capture profiler traces.

Yes, the diffusion benchmark and hotspot profiling workflow is applicable across various diffusion models and backends to locate bottlenecks and verify checked-in benchmarks.

Fallbacks occur when native diffusion paths are invalid; the guarded workflow validates native paths, detects fallbacks, and maps hotspots to guide kernel optimization handoffs.

Verify checked-in benchmarks by running the SGLang native diffusion benchmark preset to capture denoise latency, torch.profiler traces, and hotspot rankings for validation.