What problem does it solve?
This Skill turns complex SGLang torch-profiler traces into a compact diagnosis of dominant kernels, overlap headroom, and likely fusion opportunities, reducing the effort required to identify performance bottlenecks.
Core Features & Use Cases
- Kernel Triage: Rank GPU kernels by stage, runtime share, category, launch count, CPU operation, and mapped Python location.
- Overlap Analysis: Compare graph-off mapping traces with formal graph-on traces to identify exposed work, hidden work, dependency risk, and practical overlap recommendations.
- Fusion Detection: Match trace patterns against source-backed SGLang fusion and overlap catalogs while distinguishing existing, disabled, unsupported, regressed, and upstream in-flight optimization paths.
- Live or Offline Workflows: Analyze existing trace files and profile directories or trigger profiling against a running SGLang server, including separate prefill and decode views.
- Use Case: When decode latency is dominated by attention preparation and MoE support kernels, use this Skill to map those kernels back to source code and determine whether an existing fusion or dual-stream path should apply.
Quick Start
Ask the Skill to triage the SGLang torch-profiler trace or profile directory at the provided path and return the kernel, overlap-opportunity, and fuse-pattern tables.
