gchanan / nccl.torch

torch bindings for nccl

nccl.torch

Torch7 FFI bindings for NVidia NCCL library.

Installation

• Install NCCL from https://github.com/NVIDIA/nccl
• Have at least Cuda 7.0
• Have libnccl.so in your library path

Collective operations supported

• allReduce
• reduce
• broadcast
• allGather

Example usage

Argument to the collective call should be a table of contiguous tensors located on the different devices. Example: perform in-place allReduce on the table of tensors:

require 'nccl'
nccl.allReduce(inputs)

where inputs is a table of contiguous tensors of the same size located on the different devices.