chrishkchris edited a comment on issue #14: rearrange contents in dist-train.md URL: https://github.com/apache/singa-doc/pull/14#issuecomment-609034750 > Then would it be better to design the APIs in this way: > > 1. For training with MPI > > ```python > # in mnist_mpi.py > if __name__ == '__main__': > sgd = ... > sgd = DistOpt(sgd) > train_mnist(sgd, sparse, topK) > ``` > > 1. For training via multiprocessing > > ```python > # in mnist_multiprocessing.py > if __name__ == '__main__': > nccl_id = ... > sgd = ... > sgd = DistOpt(sgd, num_gpu, nccl_id) > train_mnist(sgd, sparse, topK) > ``` > > Even if you use socket, multiprocessing can only run on a single node, hence num_gpu = gpu_per_node. Sorry, I updated my comment: num_gpu is the local rank of a specific process gpu_per_node is the total number of ranks in a single node see https://github.com/apache/singa/blob/master/examples/autograd/mnist_multiprocess.py#L42 ``` for gpu_num in range(0, gpu_per_node): process.append(multiprocessing.Process(target=train_mnist_cnn, args=(sgd, max_epoch, batch_size, True, data_partition, gpu_num, gpu_per_node, nccl_id))) ``` The details of the terms: (ii) gpu_per_node gpu_per_node is the number of GPUs in a node you would like to use for training. For single node training, it is the world size. (iii) gpu_num gpu_num determine the local rank of the distributed training and which gpu is used in the process. We used a for loop to run the train function where the argument gpu_num iterates from 0 to gpu_per_node. In this case, different processes can use different GPUs for training.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
