Inferlab Integration for SGLang

Framework-specific planning and rendering for running SGLang servers through Inferlab. The consuming workspace supplies SGLang and its hardware runtime; this package supplies only the Inferlab integration boundary.

See the Inferlab repository for workspace authoring and supported topology documentation.

When profiling intent is enabled, the integration declares every single, prefill, and decode model-serving replica as a capture target of the effective mechanism. Managed collection opens the Nsight Systems range through POST /start_profile with SGLang's CUDA_PROFILER activity and closes it through POST /stop_profile; engine trace renders the control-plane-assigned trace directory into SGLANG_TORCH_PROFILER_DIR and controls the window through the same action pair with the GPU activity. InferLab, not this package, owns the profiler lifecycle, capture plan, report verification, cleanup, and records.

For direct single, setting enable_metrics = true enables SGLang's native Prometheus endpoint and declares /metrics to InferLab. Without that effective setting the integration advertises no server-metrics capability.