What problem does it solve? Recommendation models running on Ascend NPUs often fail to meet latency or throughput targets, and engineers struggle to interpret MindStudio Profiler output to find bottlenecks. This Skill parses profiling data and optional dump graphs to pinpoint H2D, ModelExecute, and D2H overhead, identify repeated subgraph structures, and produce actionable fusion and scheduling optimization suggestions. ## Core Features & Use Cases - Profiling Analysis: Parses MindStudio Profiler output (op_summary, op_statistic, api_statistic, step_trace, msprof JSON) to compute H2D/ModelExecute/D2H time, iteration counts, dynamic-shape detection, and top time-consuming operators. - Graph Structure Analysis: Analyzes GE Build pbtxt graphs and PyTorch fxgraph output_code files to find top repeated subgraphs and recommend fusion passes or fused operators. - Verified Optimization Suggestions: Generates prioritized recommendations covering AutoFuse, multi-instance parallelism, AICore control, CPU affinity, and multi-stream scheduling, with strict rules requiring all parameters to be verifiable against official Ascend documentation. - Use Case: A customer's recommendation inference service misses its latency SLA on Ascend NPU. Provide the MindStudio Profiler directory and the GE Build pbtxt; the Skill produces a Markdown report identifying that vector operators dominate and AutoFuse is disabled, then suggests enabling it with the correct environment variables. ## Quick Start Analyze the MindStudio Profiler output directory ./profiling_output for my recommendation model and generate a performance bottleneck report with optimization suggestions.
