sglang-diffusion-benchmark-profile

You can profile diffusion denoise latency by running standardized benchmark presets to collect end-to-end latency and capturing torch.profiler traces to rank dominant CUDA kernels in SGLang.

The best way to identify existing fast paths is to map profiler hotspots against packed QKV, normalization, RoPE fusions, and distributed overlap patterns to see if an optimization already exists.

Yes, CUDA-capable GPUs and PyTorch profiling support are required for benchmarking, while optional Hugging Face authentication is only needed for accessing gated models during the workload.

Yes, you can compare a tuned run against a baseline using structured perf dumps to determine if a slowdown is caused by configuration, a missed existing optimization, or a new kernel opportunity.

Yes, SGLang benchmarking supports both image and video generation workloads by running standardized presets across models like Wan, FLUX, Qwen, and LTX-2 to measure denoise latency and peak GPU memory.

High denoise latency after kernel optimizations may persist if the bottleneck stems from configuration or a missed fast path, diagnosable by classifying hotspots against existing fused kernels.