What problem does it solve? It turns raw Ascend NPU profiling output (kernel_details.csv, trace_view.json, step_trace_time.csv) into actionable bottleneck diagnoses, answering why a model is slow, where time goes, and whether an optimization actually helped. ## Core Features & Use Cases - Five-layer analysis pipeline: warmup validation, DiT/VAE stage separation, operator category breakdown (FA/MatMul/Vector/Comm), host-bound and communication-overlap analysis, and P0-P2 prioritized optimization directions via analyze_trace.py. - Baseline vs optimized comparison: compare_traces.py diffs two profiling runs at kernel level, flagging REGRESSION or improvement with an automatic PASS/WARN/FAIL verdict. - Fusion candidate identification: detects fusion opportunity candidates from kernel execution sequences and hands them to fusion-scope-analyze for boundary and benefit judgment. - Use Case: After collecting profiling data from a remote NPU, run the pipeline to learn that DiT MatMul exceeds 50% of step time, then receive a P1 recommendation pointing to MatMul quantization documentation. ## Quick Start Ask the assistant to analyze the profiling output directory with analyze_trace.py and explain where the inference bottleneck is and which optimization direction to try first.
