MusaCoder: Native GPU Kernel Generation with Full-Stack Training on Moore Threads GPU
Researchers have developed MusaCoder, a novel framework for generating native GPU kernels, which are essential for efficient low-level code execution. This system employs a full-stack training approach, integrating data synthesis, rejection fine-tuning, and reinforcement learning with a specialized verification environment called MooreEval. MusaCoder introduces several techniques to stabilize the reinforcement learning process, leading to improved correctness and speedup compared to existing models. The framework demonstrates strong performance, with its larger version setting a new state-of-the-art for native GPU kernel generation. AI
IMPACT Establishes a new state-of-the-art in native GPU kernel generation, potentially accelerating AI development on emerging hardware.