Methods

Available agent methods in KernelGenBench.

Overview

Method	Description	Model Support
naive_cc	Single Claude Code call	Opus, GLM, Qwen, MiniMax
normal_cc	Claude Code + self-verification loop	Opus, GLM, Qwen, MiniMax
naive_opencode	Single OpenCode call	Opus, GLM, Qwen, MiniMax
normal_opencode	OpenCode + self-verification loop	Opus, GLM, Qwen, MiniMax
AutoKernel	Automated kernel optimization (standalone: test_autokernel.sh)	GLM, Qwen
AKO4ALL	Kernel optimization for all operators (standalone: test_ako4all.sh)	Opus
cuda-optimized-skill	CUDA optimization with strategy memory (standalone)	Opus

Basic Agent Frameworks

Claude Code

Variant	Description
naive_cc	Single generation process
normal_cc	Generation + self-verification + iterative repair

Best for: High accuracy requirements

OpenCode

Variant	Description
naive_opencode	Single generation process
normal_opencode	Generation + self-verification loop

Best for: Alternative agent framework

Kernel-Specific Agents

AutoKernel

Automated kernel optimization pipeline:

• Focus on performance optimization

• Support for multiple models (GLM-5.0, Qwen3.5)

• Highest speedup results

AKO4ALL

Kernel optimization for all operators:

• Uses Opus-4.6

• Comprehensive operator coverage

• Good balance of accuracy and speedup

cuda-optimized-skill

CUDA optimization with strategy memory:

• Leverages historical optimization strategies

• Specialized for CUDA kernel optimization

• Strategy-based approach for improved performance

Choosing a Method

Goal	Recommended Method
High accuracy	normal_cc (Claude Code)
High speedup	AutoKernel (Qwen)
Balanced performance	AKO4ALL
Cost-effective	naive_cc