Comment by insumanth

Anyone not doing distributed training across thousands of GPU, pytorch is the best option. Others have an infra team to work on it.