sglang-diffusion-performance

To optimize SGLang diffusion performance under VRAM constraints, benchmark native pipeline configurations across your models and GPUs to select appropriate flags like --backend and offload options. This identifies the fastest configuration that fits your memory budget.

Benchmarking diffusion throughput in SGLang multi-GPU environments involves running native diffusion benchmark scripts and collecting perf-dumps. Interpreting these results guides the selection of parallelism strategies and flags to maximize throughput.

Enabling torch compile can improve SGLang diffusion speed. Use the --enable-torch-compile flag during benchmarking to evaluate its impact on your specific model and GPU configuration, alongside --warmup and precision options.

You can tune Wan2.2 or LTX-2 deployments by using the diffusion performance tuning workflow. This evaluates different backends and offload options to quickly identify the fastest configuration meeting your VRAM budget.

SGLang diffusion requires warmup to allocate memory and compile kernels before achieving peak performance. Applying the --warmup flag ensures the pipeline is fully initialized, allowing benchmark scripts to measure stable, maximum throughput accurately.