AMDKernelVault is a new system designed to accelerate GPU kernel development for AMD CDNA GPUs. The project focuses on creating a large corpus of optimized kernels, primarily using HIP and Triton. The system utilizes agent-driven pipelines, specifically HIPKernelGen and TritonKernelGen, to automatically transform PyTorch references into optimized kernels. These agents compile and validate candidate kernels under the ROCm environment and then profile their latency on AMD hardware. The resulting corpus contains 62,153 execution-verified HIP kernel samples, 2,377 production-grounded ROCm Libraries QA entries, and 39,893 Triton kernels.
The training framework incorporates Qwen3-8B, utilizing both supervised fine-tuning and execution-aware reinforcement learning. This approach leverages the generated kernel corpus to improve the model's performance. Under fixed evaluation budgets, Qwen3-8B achieved 34.0% Pass@1 on PyTorch-to-HIP, 33.2% Corr@3 on TritonBench-G, and 41.94% Corr@3 on ROCmBench. These results demonstrate the effectiveness of the AMDKernelVault corpus in improving model correctness.
The project’s availability is key to its potential. The corpus, documentation, and associated training and kernel-generation code are accessible via provided URLs. This open-source approach allows for broader experimentation and further optimization within the AMD ecosystem. The system aims to reduce the manual effort involved in GPU kernel development, particularly for complex workloads.
The development of AMDKernelVault addresses a gap in existing LLM-based kernel agents, which are often focused on CUDA/NVIDIA environments. This new framework provides a dedicated solution for AMD GPU optimization, utilizing agentic training and a substantial kernel corpus. Source: https://arxiv.org/abs/2609.12471
