What problem does it solve? Evaluating operator-level accuracy and performance on MetaX (MACA) GPUs requires coordinating container setup, environment verification, compilation, baseline generation, and result collection. This Skill automates that entire workflow so engineers can produce consistent, contract-compliant benchmark results without manual orchestration. ## Core Features & Use Cases - Operator Accuracy Validation: Verifies MetaX forward outputs, input gradients, and parameter gradients against image-bundled CPU ground truth, producing CSV/JSON pass results. - Performance Benchmarking: Runs GEMM and Conv2d (FP16/FP32), LongTail-Bench operators, and Transformer encoder/decoder block latency tests, writing measured baselines in milliseconds. - Result Contract 2.0 Collection: Uses runner and collector scripts to validate workload fingerprints, case identities, and latency values, then atomically generates a compliant result.json with summary statistics (avg, P50, P95, min, max). - Use Case: An engineer needs to validate a new MetaX driver release. They trigger the Skill to launch the benchmark container, run all GEMM and Conv2d cases, and receive a verified result.json with per-case latencies and accuracy pass rates. ## Quick Start Ask the agent to run the MetaX operator benchmark for GEMM and Conv2d on the available MetaX GPU and collect the results into result.json.
