Cody Yu, Amazon Web Services
Apache TVM: An End-to-End Deep Learning Compiler Framework for All Accelerators
As deep learning compilers are getting more and more attention due to its flexibility of supporting the latest deep learning models and ASIC accelerators, Apache TVM grows rapidly since its first announcement in 2018. In this talk, I will introduce Apache TVM and its current status. For example, Apache TVM is capable of compiling deep learning models from various frameworks, such as TensorFlow, Keras, PyTorch, MXNet, and ONNX, for various hardware devices, including CPU, GPU, FPGA, and ASICs. It has auto-tuning frameworks to optimize model inference performance on not only cloud but edge platforms. It also has a friendly interface to integrate ASIC accelerators for heterogeneous execution. Meanwhile, I will also briefly introduce the open-source community of Apache TVM and how to participate. Roughly, the TVM community merges 3-7 pull requests per day, contributed by 37 committers and 500+ contributors. The public discuss forum has more than 2K user visits each month.
Cody is an applied scientist at AWS AI as well as a committer of Apache TVM. His current main focus is the development and optimization of deep learning compilers. Most projects he involved have paper publication and already publicly open source, such as TVM auto scheduler and bring your own codegen. Meanwhile, he also served as a committee member and a reviewer of several internal conferences and journals, such as VLSI-DAT, TCAD, and DAC. Prior to joining to AWS, Cody received his master at National Tsing Hua University and PhD at UCLA in 2013 and 2019, respectively.