Skip to content
Projects

ornith-ai/CUDA-L1 has 328 GitHub stars and 111 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-19 03:22 UTC. Last push 2025-11-03.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning

Learning

README.md

Original
CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning

License     |     🏠  Project Page     |     📄  Paper    

CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning

🥳 Introduction

In this paper, we introduce CUDA-L1, an automated reinforcement learning (RL) framework for CUDA optimization. The core of CUDA-L1 is a contrastive RL model, a newly-designed RL system to enhance optimization through comparative learning.

Evaluation Results

Fig:Average speedup across different architectures on KernelBench over baselines.

🗒️ To-do List

  • Fix KernelBench evaluations with proper stream timing synchronization ✅
  • Remove caching ✅
  • Compare with torch.compile ✅
  • Compare with pytorch eager + cuda graph ✅
  • Compare with custom torch CUDA/cuDNN backend flags ✅
  • 5090/4090

🩺 Evaluation Results

Our evaluation is conducted on the KernelBench dataset, a collection of 250 PyTorch workloads designed to evaluate language models' ability to generate efficient GPU kernels.

Table: Performance comparison across different configurations on KernelBench on A100.

ConfigurationMethodMeanMax75%50%25%Success↑
# out of total
Speedup↑
>1.01x out of total
DefaultAll3.12×120×2.25×1.42×1.17×249/250226/250
Torch CompileAll2.77×69.0×2.55×1.72×1.14×249/250203/250
Torch Compile ROAll2.88×80.1×2.48×1.67×1.13×249/250200/250
CUDA GraphAll2.81×97.9×1.83×1.20×0.954×249/250147/229
• RO = Reduce Overhead
• Success and Speedup indicate the number of successful benchmarks out of the total for each level

Table: Mean speedup across different configurations and GPU devices.

ConfigurationA1003090H100H20L40
Default3.12×2.51×3.85×2.38×3.13×
Torch Compile2.77×2.58×2.74×2.89×2.85×
Torch Compile RO2.88×2.61×2.77×2.82×2.89×
CUDA Graph2.81×3.34×2.23×2.20×3.98×

❓ How to reproduce the results?

We provide CUDA code snippets optimized by CUDA-L1 in the optimized_cuda_code folder, with separate versions for each GPU device. For example, to reproduce our results on H100 XSM, download ./optimized_cuda_code/h100.json and run each code snippet on your H100 device.

📁 Structure of Release Code

Each line in the release file contains a JSON object with the following fields:

Field Description
level_id Level index in KernelBench (values: 1, 2, 3)
task_id Task index for that level
ref_code Reference CUDA code provided by KernelBench
custom_code Optimized code generated by CUDA-L1
cuda_graph_code KernelBench reference code with CUDA Graph modifications
score_default Execution time ratio: ref_code / custom_code
score_torch_compile_default Execution time ratio: ref_code / custom_code (with torch.compile)
score_torch_compile_reduce_overhead Execution time ratio: ref_code / custom_code (with torch.compile reduce_overhead mode)
score_cuda_graph Execution time ratio: cuda_graph_code / custom_code

Note: If custom_code is None, it means the RL either failed to generate code faster than the reference code or simply copied the reference code during generation.

Example Entry Structure

{
  "level_id": 1,
  "task_id": 1,
  "ref_code": "import torch...",
  "custom_code": "import torch...",
  "cuda_graph_code": "import torch...",
  "score_default": 1.762,
  "score_torch_compile_default": 1.958,
  "score_torch_compile_reduce_overhead": 2.118,
  "score_cuda_graph": 1.566,
}

🔭 Limitations and Challenges

During the training process, we found that RL is particularly susceptible to reward hacking. We've already identified quite a few hacking cases (e.g., exploiting timing measurements & caching results). If you identify any additional reward hacks in the code, we would greatly appreciate you letting us know.

📇 Citation

@article{deepreinforce2025cudal1,
  title={CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning},
  author={Li, Xiaoya and Sun, Xiaofei and Wang, Albert and Li, Jiwei and Chris, Shum},
  journal={arXiv preprint arXiv:2507.14111},
  year={2025}
}

✉️ Contact

If you have any questions, please reach out to us at [email protected].