Training and serving large-scale neural networks with auto parallelization.

A GPU-accelerated DNN inference serving system that supports instant kernel preemption and biased concurrent execution in GPU scheduling.