virtaitech/orion-client repository overview

⁠Overview

This is the official Docker Hub Registry for Orion AI Platform, developed by VirtAI Technologies Company Limited⁠.

For an overview of Orion AI Platform, please visit our GitHub repository⁠.

⁠What's New

• Nov 20, 2019 Make SHM config easy to use and support more libraries and DL frameworks

  We no longer need to mount /dev/shm/orionsock into container!

  Launch container with --ipc host and shared memory acceleration works like a charm.

  • docker run: with --ipc host
  • Kubernetes: set hostIPC: true in client yaml

  Support TF 2.0, PyTorch 1.3, and NVCaffe

  Support NCCL 2.4.x

• Oct 29, 2019 Provide PaddlePaddle 1.5 container image

• Oct 25, 2019 k8s-based deployment of Orion vGPU components

• Sep 3, 2019 Orion vGPU software supports multiple CUDA versions Orion Server supports multiple CUDA versions, as long as the CUDA SDKs are placed uner /usr/local, e.g., /usr/local/cuda-10.0, /usr/local/cuda-9.0

  It is super clean and easy to use: users DO NOT need to create symbol links at /usr/local/cuda, nor need to set environment variables.

  For Orion Client, users need to install the Orion Client runtime corresponding to a specific CUDA version.

  Please visit our github repository⁠ to access the latest Orion Server, Orion Controller, and multiple versions of Orion Client installer.

⁠Orion Client Images⁠

All the docker images are built with MLNX_OFED 4.5-1.0.1.0 user-mode driver.

⁠TensorFlow Images with RDMA Support

⁠virtaitech/orion-client:cu10.0-tf2.0-py3⁠

TF 2.0 is build from source so as to ensure that CUDA library is dynamically linked.

⁠virtaitech/orion-client:cu10.0-tf1.13-py3⁠

TF 1.13 is build from source so as to ensure that CUDA library is dynamically linked.

⁠virtaitech/orion-client:cu9.0-tf1.12-py3⁠

TF 1.12 is installed using official pip wheel.

⁠PyTorch Images with RDMA Support

⁠virtaitech/orion-client:cu10.0-torch1.3.0-py3⁠

We build PyTorch 1.3.0 and TorchVision 0.4.2 from source so as to ensure that CUDA and NCCL are dynamically linked.

⁠virtaitech/orion-client:cu9.0-torch1.1.0-py3⁠

We build PyTorch 1.1.0 and TorchVision 0.3.0 from source so as to ensure that CUDA and NCCL are dynamically linked.

• PyTorch examples⁠, git cloned into /root/examples

We also git cloned PyTorch examples⁠ into /root/examples and placed preprocessed MNIST dataset in /root/examples/data/MNIST. As a result, users can pull image and starts running MNIST example immediately.

See the repository⁠ (in Chinese) for more details on the compilation options used to build PyTorch 1.1.0 from source.

⁠PaddlePaddle Images with RDMA Support

⁠virtaitech/orion-client:cu10.0-paddle1.5-py3⁠

We compiled PaddlePaddle 1.5 from source.