What problem does it solve? NPU profiling output (kernel_details.csv) is a flat list of thousands of kernels with no model structure attached, making it hard to see which attention/FFN/MoE layer is slow or jittery. This Skill slices the kernel sequence into component instances based on the user's model description, buckets operators into semantic clusters, and produces a single-page HTML report of wall/bubble time medians and outlier layers. ## Core Features & Use Cases - Structure-aware breakdown: Splits a single-step kernel sequence into component instances (attn/ffn/moe/mtp) using user-acked stream samples, with hard warnings that must be reviewed before proceeding. - Cluster statistics and jitter detection: Groups operators inside each component by user-defined rules, computes wall_ms/bubble_ms medians, detects outlier layers, and renders a single-page HTML report plus per-component CSV/trace splits. - Incremental reuse and run comparison: Reuses structure and cluster specs across profiling runs via a small/medium/large change classification, and compares multiple runs in a history HTML table with delta percentages. - Use Case: Given a decode-phase profile of an MoE model, identify that layer 17's MoE cluster has abnormally high wall time and that a specific auxiliary stream is temporally displaced, then drill into that layer's sliced CSV for root-cause analysis. ## Quick Start Ask the agent to break down your NPU profiling results by providing the kernel_details.csv and trace_view.json paths, your model script files, and a description of the model's layer structure.
