KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation
KernelZero-7B beats Claude-4.5-Sonnet and DeepSeek-V4-Pro on CUDA and Triton kernel generation.
KernelZero co-evolves a Proposer that generates Torch modules and a Coder that translates them into CUDA or Triton kernels. Frontier-driven module generation supplies training data aligned with the Coder's weaknesses, and Correctness-Aware Group Relative Policy Optimization improves performance only after correctness is reliable. KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton. On KernelBench Levels 1 and 2, CUDA pass@1 is 75.8% and 69.6%, with pass@10 of 100% and 97%; Triton pass@1 is 77.2% and 72.5%.
- A Proposer writes Torch modules and a Coder emits CUDA or Triton kernels.
- Frontier-driven generation targets the Coder's current weaknesses.
- CA-GRPO optimizes speed only after correctness is reliable.
- CUDA pass@1 is 75.8% and 69.6% on KernelBench Levels 1 and 2.
- KernelZero-7B beats Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton.
- Products
- CUDATritonKernelBench
From the original article
1 min read at huggingface.coHigh-performance GPU kernels are essential to modern machine learning systems, yet automatically generating kernels that are both correct and efficient remains challenging. Existing LLM-based approaches face two major limitations: the scarcity of high-quality training data aligned with the model's current capabilities, and the inherent trade-off between kernel correctness and performance.
By Changxin Ke, Rui Zhang, Zixiang Fang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.