ascend-moe-optimizer-trace-analyzer

What problem does it solve?

It helps you quickly identify performance bottlenecks in Ascend Mixture-of-Experts (MoE) runs by turning Chrome/Perfetto trace.json data into structured phase/category/core-group statistics and an interpretable diagnosis, instead of manually inspecting timelines.

Core Features & Use Cases

• Trace-to-Stats Pipeline: Parses trace.json (including X events and B/E paired events) and maps raw event names into deterministic phases via a configurable phase map.
• Bottleneck-Ready Reporting: Produces CSV and Markdown reports with phase instances, summaries by phase/category/core-group, overlap analysis, and bubble (unattributed time) detection.
• Deterministic Auto Diagnosis (Optional LLM Expansion): Generates structured, evidence-based findings and can optionally append an LLM-produced expert interpretation.
• MoE/FusedDeepMoe Focus: Comes with a default phase mapping optimized for UMDK-style FusedDeepMoe traces, while still supporting custom trace mappings when analyzing other sources.

Quick Start

Use the ascend-moe-optimizer-trace-analyzer skill to analyze your trace.json file by running python3 app.py --trace trace.json --output-dir output/case1.