ascend-profiling-anomaly

To detect device idle gaps and bubble anomalies in Ascend NPU profiling data, analyze kernel_details.csv and trace_view.json files to reveal multi-stream timing issues and host-device delays. This process identifies wait-pollution risks and attributes soft root-causes.

Reverse-engineering model architecture from Ascend profiling traces involves parsing op_summary and communication.json data to extract layer structure and per-layer kernel composition. This yields a separate Markdown report detailing the model architecture insights.

Yes, you can analyze host-device delays and wait-pollution risks by applying anomaly detection across op_summary and communication.json files. This identifies multi-stream timing bubbles and produces a structured anomaly report for Ascend NPU captures.

Profiling anomaly detection requires specific formats including kernel_details.csv, trace_view.json, op_summary, and communication.json. Applying the analysis across these profiling captures reveals device idle gaps, host-device issues, and architecture insights.

Generate a structured anomaly report and architecture summary by running profiling data analysis across kernel_details.csv, trace_view.json, op_summary, and communication.json. This returns standalone deliverables covering anomaly discovery and model architecture.