NVIDIA/torch-nccl

Name: torch-nccl

Owner: NVIDIA Corporation

Description: torch bindings for nccl

Forked from: ngimel/nccl.torch

Created: 2016-08-15 23:09:11.0

Updated: 2016-08-16 15:12:35.0

Pushed: 2016-06-23 16:32:03.0

Homepage: null

Size: 103

Language: Lua

GitHub Committers

User	Most Recent Commit	# Commits

Other Committers

User	Email	Most Recent Commit	# Commits

README

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:

ire 'nccl'
.allReduce(inputs)

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