sglang-diffusion-benchmark-profile

To profile diffusion latency, run a denoise benchmark on your target model to capture a perf dump. This identifies slow operations and ranks hotspots to guide your kernel optimizations.

Yes, you can collect torch.profiler traces to diagnose diffusion bottlenecks. The profiling workflow supports capturing detailed traces and mapping identified hotspots to known fast paths for optimization.

Benchmarking denoise latency involves running preflight checks and capturing perf dumps across various models, presets, and hardware. This process identifies slow operations and verifies hotspots before starting kernel work.

You can verify improvements by running before and after comparisons using the benchmarking and profiling workflow. This compares denoise latency and torch.profiler traces to confirm your kernel optimizations resolved the bottlenecks.

Yes, after mapping hot kernels to known fast paths and fusion families, the workflow routes confirmed issues directly to kernel-optimization skills to streamline your diffusion backend optimizations.