sglang-diffusion-benchmark-profile

What problem does it solve?

Diffusion model benchmarks and profiling streamline the process of measuring denoise latency, end-to-end generation timing, and memory usage, enabling data-driven optimization decisions.

Core Features & Use Cases

• End-to-end diffusion benchmarks: measure denoise latency, end-to-end latency, and peak memory across presets.
• Profiling workflow: collect torch.profiler traces and perf dumps to rank hot kernels.
• Guided optimization handoff: map hotspots to existing fast paths and hand off kernel work to specialized optimization skills when needed.

Quick Start

Run the diffusion benchmark profile for a chosen model to measure denoise latency and collect profiling data.