evokernel 0.2.0

CUDA kernel optimization agent powered by OpenEvolve - fork of gptme

EvoKernel

CUDA kernel optimization agent powered by OpenEvolve.

Quick Start

# Install as CLI tool (requires Python 3.10-3.13)
pipx install evokernel
# or with uv
uv tool install evokernel

# Setup Modal (GPU evaluator)
modal setup

# Configure OpenRouter API key
evokernel setup

# Run in any project directory
evokernel

Development Install

git clone https://github.com/haladir-ai/EvoKernel.git
cd EvoKernel
uv tool install .

Usage

Just run evokernel in a directory with CUDA kernels and chat with the agent:

> optimize my_kernel.cu

> check status

> show the best result

> apply changes

Current Features

• Speedup optimization: Maximizes kernel throughput via evolutionary search
• EVOLVE markers: Define which code regions to optimize
• Modal GPU evaluation: Compile and benchmark on cloud A100 (default)
• Local GPU evaluation: Run benchmarks on your own GPU (for codebases with dependencies)
• LLM ensemble: Uses Gemini, Claude, GPT via OpenRouter

Evaluator Modes

Mode	Use Case
modal (default)	Self-contained kernels, no local GPU needed
local	Kernels with custom headers/dependencies

Roadmap (Not Yet Implemented)

• Memory optimization goal: Minimize memory usage instead of speedup
• Target hardware selection: Optimize for specific GPUs (A100, H100, V100)
• Energy efficiency goal: Minimize power consumption