sglang-torch-profiler-analysis

To analyze SGLang torch-profiler traces for overlap opportunities, you can use a unified analysis workflow that maps GPU kernels to Python frames and CPU ops. It reads trace JSONs, builds kernel maps, and outputs compact tables plus an ASCII timeline to reveal actionable deployment insights.

Attributing GPU kernels to Python frames in a torch-profiler trace involves reading the trace JSON and building a kernel map. This process recovers Python locations for kernels, enabling detailed source attribution and exposing actionable deployment insights for SGLang servers.

Yes, torch-profiler trace analysis is applicable to both tracing from running SGLang servers and from stored trace files or directories. You can analyze existing trace JSONs directly without needing a live server connection.

Two-trace triage for SGLang profiler overlap analysis uses a mapping trace plus a formal trace for real overlap assessment. This workflow produces a unified three-table triage: kernel mapping, overlap opportunities, and fuse opportunities.

Yes, the torch-profiler analysis workflow includes Perfetto-friendly post-processing with optional trace repair to improve Perfetto rendering. This ensures the generated trace JSONs are compatible and display correctly within the Perfetto UI.

To analyze SGLang torch-profiler traces, you need the included Python scripts and the proper server configuration. The analysis reads trace JSONs from either a running SGLang server or stored directories to build kernel maps and output tables.