# BytedTsinghua-SIA/CUDA-Agent

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

Repository: https://github.com/BytedTsinghua-SIA/CUDA-Agent
Canonical: https://ross.abutalabs.com/products/cuda-agent
Homepage: https://cuda-agent.github.io/
Language: Python
License Family: other
Last push: 2026-07-08T12:01:09+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 91, release rhythm 35, longevity 15
- inputs: {"age_days": 212, "days_push": 56, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1256, forks 121 (observed 2026-08-28T04:04:08.980771+00:00)

## What it is
CUDA-Agent is a large-scale agentic reinforcement learning system from ByteDance Seed and Tsinghua that trains LLMs to generate high-performance CUDA kernels, achieving state-of-the-art results on KernelBench and outperforming torch.compile. It releases the CUDA-Agent-Ops-6K training dataset, an expert-designed SKILL.md agent environment, and a standardized workspace for the implement-compile-verify-profile loop.

## Use cases
- generate optimized CUDA kernels with an LLM agent
- train a model with agentic RL for GPU kernel code generation
- benchmark CUDA kernel generation against torch.compile on KernelBench
- synthesize training data for CUDA operator fusion tasks
- set up an agent workspace to compile, verify, and profile custom CUDA kernels
- speed up PyTorch models with RL-generated CUDA extensions

## When to choose
- you need state-of-the-art CUDA kernel generation that beats torch.compile
- you want to reproduce or extend agentic RL training for code optimization
- you need a curated dataset of executable, deterministic CUDA operator tasks
- you want a ready-made environment for the kernel implement-verify-profile loop

## When to avoid
- you need a production inference library rather than a research training system
- you have no NVIDIA GPU or CUDA development toolchain
- you want a no-code tool for accelerating PyTorch models
- you require a permissively licensed dependency - the repo has no license file

## Facets
- artifact type: framework
- maturity: active
- function: machine-learning, llm-training, agent-framework, gpu-computing, benchmarking, data-generation
- domain: machine-learning, gpu-computing, artificial-intelligence, reinforcement-learning, large-language-models, deep-learning, performance
- platform: python
- tags: cuda-kernels, reinforcement-learning, agentic-rl, kernel-generation, kernelbench, pytorch, code-generation, llm-agents, training-dataset, gpu-optimization, ai-agents, gpu, linux, docker

## Member repositories
- BytedTsinghua-SIA/CUDA-Agent (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:08.980771+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T05:07:14.177913+00:00, confidence not recorded.
  - readme: https://github.com/BytedTsinghua-SIA/CUDA-Agent (fetched 2026-08-28T04:04:08.980771+00:00, sha be4bb0ca30e6)
  - homepage: https://cuda-agent.github.io/ (fetched 2026-08-29T12:17:47.989753+00:00, sha 24a2b6181eb7)
- Data as of 2026-08-30T08:39:29.467469+00:00.