What problem does it solve? Measuring NLP operator performance on Hygon DCU accelerators requires manual container setup, environment validation, benchmark execution, and result formatting, which is error-prone and hard to reproduce. This Skill automates the full evaluation pipeline so operators get consistent, verifiable latency baselines. ## Core Features & Use Cases - Multi-operator benchmarking: Runs GEMM (FP16/FP32 via GPUfusion), Conv2d (HIP-PyTorch/MIOpen), LongTail-Bench operators, and Transformer encoder/decoder block inference latency tests. - Deterministic result collection: Converts measurement CSVs into Result Contract 2.0 JSON with per-case latency and eight summary metrics (avg, p50, p95, min, max), with run-token and workload-fingerprint validation to prevent stale result reuse. - Containerized workflow: Guides Docker container creation with DCU device mounts, DTK environment checks, and atomic result publishing. - Use Case: An engineer needs baseline GEMM latency on a Hygon DCU cluster. The Skill launches the container, runs FP16/FP32 GEMM with 10 warmup and 1000 timed iterations, and emits a validated result.json. ## Quick Start Ask the agent to generate GEMM operator baseline values on a Hygon DCU and it will run the full benchmark and produce result.json.
