gpu-inference

Tesla P100's silent FP16 precision bug in llama.cpp fixed

A 3-line CUDA patch in llama.cpp fixes a years-old silent FP16 precision bug on Tesla P100 GPUs, with zero performance cost.