There has been a blossoming open source scene for deep learning frameworks, such as PyTorch, TensorFlow, and MXNet. However, the democratization of AI will only truly be achieved when the cloud infrastructure running these frameworks is open and better optimized.
Currently, various intermediate representation (IR) solutions like NNVM, DLVM, XLA, and nGraph help bridge the gap between the framework frontend and the hardware backend. But in order for the cloud platforms to truly be perceptive of the AI workloads, resource representation (RR) is needed to provide a mapping relationship between the resource provided by the infrastructure (GPU, FPGA, ASIC, etc.) and a given IR, which requires the resource to do the compiling.
Zhipeng Huang explains how resource representation works with various intermediate representation technologies to help achieve the democratization of AI. Zhipeng begins with a brief overview of the motivation as well as the current open source IR solutions, then provides a deep dive into how cyborg projects could help user build a open source cloud infra via OpenStack or Kubernetes for its AI application via RR techniques.
Zhipeng Huang is an open source operation manager for Huawei. Zhipeng has been involved with various major open source communities and is now the PTL of OpenStack Cyborg project, cochair of OpenStack Public Cloud WG, and co-lead of the Kubernetes Policy WG. His current focus is open source cloud federation (OpenStack Tricircle, Ubernetes, Mesos Federation), and storage function visualization with open source software and hardware solutions.
©2018, O’Reilly UK Ltd • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • firstname.lastname@example.org