What problem does it solve? Measuring NLP operator performance on Iluvatar BI-V150 GPUs requires manual container setup, CoreX toolchain compilation, benchmark execution, and result formatting, which is error-prone and hard to reproduce. This Skill automates the full evaluation pipeline and produces standardized Result Contract 2.0 output. ## Core Features & Use Cases - Multi-Operator Benchmarking: Runs GEMM (FP16/FP32 via CoreX CUDA), Conv2d (PyTorch 2.7.1/CoreX backend), LongTail-Bench operators, and Transformer encoder/decoder block inference latency tests. - Deterministic Runners and Collectors: Uses dedicated scripts to compile kernels, execute benchmarks with run-scoped tokens, and convert CSV results into validated result.json with summary metrics (avg, p50, p95, min, max latency). - Containerized Workflow: Guides Docker container creation with the prebuilt CoreX 4.4.0 image, volume mounts for project code, results, and logs, plus environment verification steps. - Use Case: An engineer needs baseline latency numbers for GEMM kernels on a BI-V150 card. The Skill starts the container, compiles the CUDA ops with CoreX, runs the FP16/FP32 case CSVs, and emits a validated result.json with per-case and summary latency metrics. ## Quick Start Ask the agent to run the GEMM operator benchmark on the Iluvatar BI-V150 GPU and generate the baseline result.json.
