tf.contrib.nccl.all_max

tf.contrib.nccl.all_max(tensors)

Defined in tensorflow/contrib/nccl/python/ops/nccl_ops.py.

Returns a list of tensors with the all-reduce max across tensors.

The computation is done with an all-reduce operation, so if only some of the returned tensors are evaluated then the computation will hang.

Args:

  • tensors: The input tensors across which to reduce; must be assigned to GPU devices.

Returns:

List of tensors, each with the maximum of the input tensors, where tensor i has the same device as tensors[i].