ascend-moe-optimizer-trace-analyzer

You can diagnose MoE execution bottlenecks by parsing Chrome or Perfetto trace.json files into structured phase and core-group timing statistics. The tool maps raw trace events through a configurable phase map to produce CSV summaries and Markdown diagnosis reports.

Yes, automatic overlap and bubble detection is supported for Ascend FusedDeepMoe traces. The analyzer computes deterministic interval aggregation to find overlapping execution phases and identifies unattributed bubble time, outputting the results into structured CSV tables.

Generating CSV reports from Chrome trace.json involves running a Python script that extracts X events and B/E paired events. It aggregates timing data by phase, category, and core-group, outputting detailed CSV files for Mixture-of-Experts profiling.

Yes, the analyzer works with custom phase mappings for non-UMDK traces. While optimized for UMDK-style FusedDeepMoe traces by default, you can supply a custom phase_map.yaml to map raw event names into deterministic phases for other trace sources.

The best way to diagnose unattributed bubble time is using automated trace analysis that computes deterministic interval unions. It identifies gaps in execution phases within your trace.json and documents them as bubble summaries in a structured Markdown report.

No, an external LLM is optional for generating performance diagnosis. The tool produces structured evidence-based findings automatically, but can optionally append an LLM-produced expert interpretation by executing validated external prompts.