The NVIDIA container image of Torch, release 17.05, is available.
Torch container image version 17.05 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
- NVIDIA CUDA 8.0.61
- NVIDIA CUDA® Deep Neural Network library™ (cuDNN) 6.0.21
- NVIDIA NCCL 1.6.1 (optimized for NVLink™ )
Key Features and Enhancements
This Torch release includes the following key features and enhancements.
- Latest cuDNN release
- Ubuntu 16.04 with April 2017 updates
Known Issues
The NCCL library version 1.6.1 included in this image, modifies the output buffers on all GPUs during in-place ncclReduce() operations, whereas normally only the "root" (target) device's output buffer should be modified. This is fixed in later versions of NCCL, as will be packaged in later versions of this image. As a workaround, either use ncclAllReduce(), which correctly modifies output buffers of all GPUs to the same values, or use out-of-place ncclReduce(), wherein the output buffer is distinct from the input buffer.