hkust-nlp

hkust-nlp / KernelGYM

Public

[KernelGYM & Dr. Kernel] A distributed GPU environment and a collection of RL training methods to support RL for Kernel Generations

94
5
100% credibility
Found Feb 06, 2026 at 42 stars 2x -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

KernelGYM is a distributed testing playground for evaluating and improving AI models that generate optimized GPU kernels, complete with ready-to-use training recipes.

How It Works

1
🔍 Discover KernelGYM

You hear about a helpful tool that tests and improves AI for writing super-fast code that runs on powerful graphics cards.

2
📥 Download easily

Grab the free tool from its trusted page with a simple download, like getting a new app.

3
⚙️ Quick setup

Run one easy command to prepare everything on your computer, no complicated steps needed.

4
🚀 Launch your gym

Click start and watch your personal testing playground come alive, ready for AI experiments.

5
🧪 Test code ideas

Send sample code to see how fast and correct it runs on the graphics card.

6
🤖 Train smarter AI

Follow ready guides to teach your AI to create even better, faster code over time.

🏆 Win with speed

Celebrate as your AI generates lightning-fast code that powers up games, science, and more!

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 42 to 94 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is KernelGYM?

KernelGYM is a Python-based distributed GPU environment for evaluating and training AI models on kernel generations. It handles GPU kernel compilation, correctness checks, and performance benchmarking across clusters, with built-in support for CUDA and Triton backends. Users get scalable parallel evaluation via API endpoints like /evaluate and /workflow/submit, plus a collection of RL training methods from the Dr.Kernel paper.

Why is it gaining traction?

It tackles real pain points in kernel eval—CUDA crashes, precise timing, and multi-node scaling—with subprocess isolation and Redis queues that keep pipelines running. Developers notice the quick single-node setup via start_all_with_monitor.sh and easy extension to custom GPU tasks like simulations. The RL integration for long-horizon training stands out for agentic optimization without reward hacking.

Who should use this?

AI researchers fine-tuning models for Triton kernel generations via RL, especially those hitting GPU eval bottlenecks. Teams optimizing kernels at scale on KernelBench datasets, or devs prototyping multi-turn agent workflows for GPU workloads. Perfect for HKUST-NLP folks or similar pushing Dr.Kernel methods.

Verdict

Grab it if you're in GPU kernel RL—solid docs, arXiv backing, and working scripts make it usable now despite 73 stars and 1.0% credibility score. Still early-stage with room for more examples and tests, but scales well for serious training.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.