ascend-moe-optimizer-auto-trace

What problem does it solve?

This Skill helps you add end-to-end TRACE_POINT instrumentation and MoE profiling capture for Ascend operators, so you can generate Chrome trace JSON that pinpoints performance bottlenecks across AIC/AIV execution paths.

Core Features & Use Cases

• Full-chain instrumentation & profiling: inserts TRACE_POINT around real operator phase boundaries (not just shell entries), captures per-core profiling tensors, and decodes them into Chrome trace events.
• Strict correctness guardrails (G1–G5): enforces trace preprocessor hook/point_map generation consistency, output arity and “profiling is last” ordering, compilation and example/UT synchronization + save steps, and absolute-path safety for spawned multiprocessing.
• Toolchain deployment + compile hook integration: can deploy trace toolchain scripts and patch existing compile scripts with a preprocessor hook for the active build tree.
• Validation workflow: runs validate_trace_points.py and check_compile_safety.py, then guides the required full compilation and profile/trace generation steps.

Quick Start

Ask for instrumentation and trace generation for an Ascend MoE operator (e.g., “Add TRACE_POINT + profiling and produce chrome_trace.json for this op”), and then follow the skill’s G1–G5 checklist, ending with running save_profiling_data after NPU synchronization and generating chrome_trace.json via trace_collector using the same build’s point_map.json.