good batching and tensor parallelization prob