Torch Release 18.02

The NVIDIA container image of Torch, release 18.02, is available.

Torch container image version 18.02 is based on Torch7.

Contents of Torch

This container image contains the complete source of the version of NVIDIA Torch in /opt/torch. It is pre-built and installed into the /usr/local/[bin,share,lib] directories in the container image.

The NVIDIA^® Collective Communications Library ™ (NCCL) (NCCL) library and Torch bindings for NCCL are installed in this container, and models using DataParallelTable can easily leverage this library for fast parallel training. The container also includes the following:

• Ubuntu 16.04 including Python 2.7 environment
• NVIDIA CUDA 9.0.176 including:
  • CUDA^® Basic Linear Algebra Subroutines library™ (cuBLAS) 9.0.282 Patch 2 which is installed by default
  • cuBLAS 9.0.234 Patch 1 as a debian file. Installing Patch 1 by issuing the dpkg -i /opt/cuda-cublas-9-0_9.0.234-1_amd64.deb command is the workaround for the known issue described below.
• NVIDIA CUDA^® Deep Neural Network library™ (cuDNN) 7.0.5
• NCCL 2.1.2 (optimized for NVLink™ )

Driver Requirements

Release 18.02 is based on CUDA 9, which requires NVIDIA Driver release 384.xx.

Key Features and Enhancements

This Torch release includes the following key features and enhancements.

• Incorporated all upstream changes.
  Note:

  Note: We will no longer incorporate upstream changes after the 18.02 release due to the lack of upstream activity.

• Latest version of cuBLAS
• Ubuntu 16.04 with January 2018 updates

Known Issues

cuBLAS 9.0.282 regresses RNN seq2seq FP16 performance for a small subset of input sizes. This issue should be fixed in the next update. As a workaround, install cuBLAS 9.0.234 Patch 1 by issuing the dpkg -i /opt/cuda-cublas-9-0_9.0.234-1_amd64.deb command.