generate-profile

To profile an SGLang server, this Skill automates the launch, validates readiness via a health endpoint, and generates a Chrome-compatible profiling trace. It handles the end-to-end workflow to produce reproducible performance data for later analysis.

End-to-end SGLang server profiling requires a working SGLang installation, at least one CUDA GPU, access to a health endpoint, and storage space for the generated trace directory. The environment must support server launch and readiness validation.

Yes, you can capture reproducible performance traces for SGLang inference. The Skill saves the trace and configuration during each profiling run, allowing you to compare latency and throughput data across multiple runs to identify bottlenecks.

The best way to automate SGLang server profiling is using an end-to-end workflow that launches the server, runs a sanity check, and captures a Chrome-compatible trace. This reduces manual profiling effort and provides trace-driven optimization data.

Yes, SGLang profiling works with a single CUDA GPU setup. The Skill requires at least one CUDA GPU to execute the profiling cycle, validate server readiness, and generate the performance trace directory for model serving optimization.