Adapting Language Models for Low-Resource GPU Kernel Programming

Published on 14 May 2026, 16:53

Speakers: Natalia Pahlavan & Laasya Konidala, Stanford University
Talk Abstract: Large Language Models perform well on high-resource programming languages, but they struggle to generate low-resource, compiler-verified languages such as AMD HIP, where open-source training data is scarce and performance constraints are strict. In this work, we investigate improving HIP kernel generation via (1) synthetic generation of 400 PyTorch tasks with full prompt histories across iterative attempts (2) multi-agent translation and optimization into HIP kernels using evolutionary search and MI350X benchmarking, and (3) SFT on the synthetic dataset followed by GRPO-based RL with hardware-in-the-loop rewards. We evaluate on KernelBench across all three levels using compilation success, functional correctness, and execution speedup relative to PyTorch baselines on the MI350X.

Find the resources you need to develop using AMD products: amd.com/en/developer.html

Join the Developer Community: devcommunity.amd.com

Join the Developer Discord server: discord.gg/amd-dev

***

© 2026 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, EPYC, ROCm, and AMD Instinct and combinations thereof are trademarks of Advanced Micro Devices, Inc.