sglang-diffusion-performance

To optimize SGLang diffusion performance, run the benchmark workflow to profile combinations of performance flags like --enable-torch-compile and offloads. This identifies the fastest configuration that meets your VRAM constraints while preserving acceptable generation quality.

Diffusion benchmarking compares native SGLang and Diffusers backends to evaluate lossless and lossy optimization paths. This comparison identifies the maximum throughput configuration for your specific model and multi-GPU setup.

Yes, you can tune diffusion latency across multi-GPU setups by profiling flag combinations such as --enable-torch-compile and --warmup. The benchmark workflow tests these flags to find the lowest latency within your fixed VRAM cap.

Yes, the SGLang diffusion benchmark supports offloads to fit VRAM constraints. You can profile and tune offloading configurations alongside other performance flags to maximize throughput without exceeding your available VRAM budget.

The best way to benchmark Wan2.2 IV on 8 GPUs is applying the SGLang diffusion benchmark workflow to profile performance flags. This process identifies the fastest configuration achieving the lowest latency within your specific hardware and VRAM limits.

Enabling torch compile can increase memory usage, causing VRAM limits to be exceeded. You can resolve this by profiling offload configurations and warmup flags through the benchmark workflow to meet your VRAM constraints.