Bruce-Lee-LY

Bruce-Lee-LY / cuda_auto_tune

Public

NCU-driven iterative optimization workflow for CUDA/CUTLASS/Triton/CuTe DSL kernels.

auto-tune claude-code codex cuda cursor
10
0
100% credibility
Found Apr 09, 2026 at 10 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This repository provides high-performance implementations of RMSNorm for AI workloads on GPUs, including benchmarks and profiling tools to iteratively optimize kernel speed using various DSLs.

How It Works

1
🔍 Discover the optimizer

You find this helpful collection of tools designed to make heavy math calculations in AI models run much faster on powerful graphics processors.

2
📥 Grab the examples

Download the ready-to-use examples and performance demos to get started right away.

3
▶️ Run the speed tests

Launch the built-in tests to see how different fast-math methods compare in timing charts.

4
📊 Spot the slowdowns

Use the smart checker to analyze exactly where your calculations are taking too long.

5
🔧 Tune for speed

Follow the simple guides to adjust your math routines based on the checker's insights.

6
⚡ Test the improvements

Re-run the checker and celebrate as your calculations now finish in record time.

🎉 Lightning-fast results

Your AI math operations are now super optimized, saving time and boosting efficiency every run.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 10 to 10 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 cuda_auto_tune?

This Python tool delivers an NCU-driven iterative optimization workflow for tuning CUDA/CUTLASS/Triton/CuTe DSL kernels. It automates deep profiling with NCU to spot bottlenecks like roofline gaps, warp stalls, and occupancy issues, then guides targeted tweaks via launch configs, memory patterns, and DSL-specific params. Developers get scripts to profile any kernel command, export CSV metrics, generate analysis reports, and re-verify improvements in a mandatory loop.

Why is it gaining traction?

Unlike ad-hoc tuning, it enforces a data-first cycle: profile, analyze multi-dimensional metrics, modify, re-profile—slashing guesswork for complex kernels. The profiling script handles native CUDA binaries, Python Triton/CuTe scripts, and even Triton cache inspection, with benchmarks showing Triton/CuTe matching FlashInfer speeds on RMSNorm for hidden sizes up to 8192. It's a practical hook for anyone chasing peak GPU throughput without manual NCU drudgery.

Who should use this?

CUDA kernel engineers optimizing ML primitives like RMSNorm in transformer models. Triton or CuTe DSL users iterating on num_warps, stages, or copy atoms. Teams profiling CUTLASS epilogues or native async copies on H100/H20-class GPUs.

Verdict

Skip for production—10 stars and 1.0% credibility signal early-stage code with thin docs and single-example focus. Grab it as a free NCU workflow template if you're manually tuning kernels; fork and expand for real wins.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.