PyTorch Release 18.05

The NVIDIA container image of PyTorch, release 18.05, is available.

Contents of PyTorch

This container image contains the complete source of the version of PyTorch in /opt/pytorch. It is pre-built and installed in the pytorch-py3.6 Conda™ environment in the container image. The container also includes the following:

• Ubuntu 16.04 including Python 3.6 environment
• NVIDIA CUDA 9.0.176 (see Errata section and 2.1) including CUDA^® Basic Linear Algebra Subroutines library™ (cuBLAS) 9.0.333 (see section 2.3.1)
• NVIDIA CUDA^® Deep Neural Network library™ (cuDNN) 7.1.2
• NCCL 2.1.15 (optimized for NVLink™ )
• Caffe2 0.8.1

Driver Requirements

Release 18.05 is based on CUDA 9, which requires NVIDIA Driver release 384.xx.

Key Features and Enhancements

This PyTorch release includes the following key features and enhancements.

• PyTorch container image version 18.05 is based on PyTorch 0.4.0.
• Includes Caffe2 0.8.1. For more information, see PyTorch and Caffe2 repos getting closer together.
• APEx, an extension providing utilities for FP16 and multi-gpu training. For more information, see APEx: A PyTorch Extension and APEx.
• Ubuntu 16.04 with April 2018 updates

Known Issues

• Some mixed-precision models might encounter a crash due to a new FP16 overflow check added in PyTorch. We have an upstream fix submitted with PR 7382 and should be resolved in a future container.
• There is a minor performance regression with the imagenet sample in /opt/pytorch/examples/imagnet for some network architectures on multi-gpu cases. This regression will be fixed in the next release.