Capabilities#

KernelGenBench provides comprehensive capabilities for Kernel generation evaluation.

Multi-Source Evaluation#

210 operators from ATen, vLLM, and cuBLAS sources.

→ See Multi-Source Operators for details.

Multi-Chip Support#

6 hardware platforms (NVIDIA, Ascend, MUSA, Hygon, Iluvatar, MetaX) with auto-detection.

→ See Multi-Chip Support for details.

Two Evaluation Tracks#

Track

Purpose

LLM Track

Pass@K evaluation

Agent Track

Iterative generation

Anti-Hack Validation#

Three-tier mechanism (AST scan, Ghost replay, Hardware profiling).

→ See Anti-Hack Architecture for details.

Evaluation Metrics#

Accuracy, speedup, token cost, wall time.

→ See Evaluation Metrics for details.