1. 프로젝트 개요

CUDA-Agent는 Claude Opus-4.6, Gemini 3 Pro와 같은 고급 모델을 고성능 CUDA 커널 생성에서 능가한 최초의 RL 학습 모델입니다. KernelBench에서 최첨단 결과를 달성하며, 난이도 수준 전반에 걸쳐 torch.compile baseline을 꾸준히 능가하며, 특히 가장 어려운 케이스에서 강한 성능 향상을 보입니다. LLM 기반 CUDA 생성 커뮤니티를 지원하기 위해, 우리는 학습 데이터, 전문가가 설계한 SKILL.md 및 에이전트 환경을 공개했습니다.

2. 데이터셋 릴리스: CUDA-Agent-Ops-6K

우리는 학습 데이터셋 CUDA-Agent-Ops-6K를 공개했습니다:

• 데이터셋 URL: BytedTsinghua-SIA/CUDA-Agent-Ops-6K
• 규모: 6,000개의 훈련 샘플
• 건설 파이프라인:
• torch 및 [[T1]에서 참조 연산자를 수집합니다.
• LLM을 사용하여 여러 연산자를 융합된 작업으로 구성하는 방법
• 실행 가능하고 결정적이며 비자명한 샘플을 유지하기 위해 규칙 기반 필터링 적용
• 필터링 기준:
• 열성 모드와 torch.compile 모두에서 올바르게 실행해야 합니다.
• 확률 연산자 제거 및 퇴화된 출력
• KernelBench 테스트와 매우 유사한 샘플 제거 및 실행 시간 제어, 오염 위험을 줄이기

3.

CUDA-Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation

1. Project Overview

CUDA-Agent is the first known RL-trained model to surpass advanced models such as Claude Opus-4.6 and Gemini 3 Pro on high-performance CUDA kernel generation. It achieves state-of-the-art results on KernelBench, consistently outperforming the torch.compile baseline across difficulty levels, with especially strong gains on the hardest cases. To support the LLM-based CUDA generation community, we have released our training data, expert-designed SKILL.md and agent environment.

2. Dataset Release: CUDA-Agent-Ops-6K

We released the training dataset CUDA-Agent-Ops-6K:

• Dataset URL: BytedTsinghua-SIA/CUDA-Agent-Ops-6K
• Scale: 6,000 training samples
• Construction pipeline:
  • Collect reference operators from torch and transformers
  • Use an LLM to compose multiple operators into fused tasks
  • Apply rule-based filtering to keep executable, deterministic, and non-trivial samples
• Filtering criteria:
  • Must execute correctly in both eager mode and torch.compile
  • Remove stochastic operators and degenerate outputs
  • Control runtime range and remove samples highly similar to KernelBench tests to reduce contamination risk

3. agent_workdir Overview

agent_workdir is a standardized agent workspace example for the full loop: implement CUDA kernels -> compile -> verify correctness -> profile performance -> iterate.

Key files in this directory:

• SKILL.md: workflow constraints and optimization rules for agent execution
• model.py: original PyTorch baseline model
• model_new.py: optimized model using the custom CUDA extension
• binding.cpp / binding_registry.h: shared Python binding registration infrastructure
• kernels/: custom CUDA/C++ kernels and their bindings
• utils/compile.py + utils/compile.sh: extension build scripts
• utils/verification.py: correctness validation script
• utils/profiling.py: performance comparison against baseline and torch.compile

Common commands (run inside agent_workdir):

bash utils/compile.sh
python3 -m utils.verification
python3 -m utils.profiling