Correct First, Fast Second: What a Local 7B Code Model Can and Cannot Do for GPU Kernel Optimization
AI-generated summary Researchers evaluated how a local 7B-parameter code model, specifically Qwen2.5-Coder-7B-Instruct, performs when replacing PyTorch operators with custom CUDA kernels across 30 KernelBench tasks on a V100 GPU. The study found that correctness remains the binding constraint, with profiling-guided prompting achieving 43 percent correct and 27 percent faster performance than the baseline.