generate-profile

To generate an SGLang server profiling trace, launch the server, validate its readiness, run an automated accuracy sanity check, and capture a Chrome-compatible trace. The process returns a profile directory containing all generated performance artifacts.

Capturing an SGLang server profile requires a working SGLang installation, at least one CUDA-capable GPU, and access to the server health endpoint and profiling utilities to successfully validate readiness and generate traces.

Yes, you can validate SGLang server performance by running an automated accuracy sanity check immediately after launching the server. This ensures the server is fully ready before you capture the Chrome-compatible profiling trace.

The SGLang profiling trace is output in a Chrome-compatible trace format. The Skill delivers a profile directory containing these trace artifacts, allowing you to verify server performance directly within compatible viewers.

No, SGLang profiling requires at least one CUDA-capable GPU to execute the server run and capture the end-to-end performance trace. The GPU is a mandatory prerequisite for generating the profiling artifacts.

Your SGLang server profile may not generate correctly if the server is not fully ready. Ensure you have access to the server health endpoint and successful completion of the accuracy sanity check before attempting to capture the profiling trace.