SLAI-AITP (AI Training Platform Team) · Kernel Optimisation Benchmark

AscendGYM

AscendGYM is a benchmark for measuring how well models and agents optimise existing Ascend NPU kernels, using 48 kernels across three difficulty levels. Test cases use data types and tensor shapes collected from real training and inference workloads. It supports model and model-harness evaluation, and introduces the Speedup-Token Score (STS), which measures speedup relative to token cost. Code is available on GitHub.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.