The Apache Software Foundation (ASF), the all-volunteer developers, stewards, and incubators of more than 350 Open Source projects and initiatives, announced today Apache® TVM™ as a Top-Level Project (TLP).

The ASF's first full stack software and hardware co-optimization project, Apache TVM is an end-to-end open deep learning compiler stack for CPUs, GPUs, and specialized accelerators. TVM enables machine learning developers to optimize and run computations efficiently on any hardware backend. The project originated in 2017 as a research project at Washington University and entered the Apache Incubator in March 2019.

"It is amazing to see how the Apache TVM community members come together and collaborate under The Apache Way," said Tianqi Chen, Vice President of Apache TVM. "Together, we are building a solution that allows machine learning engineers to optimize and run computations efficiently on any hardware backend."

Apache TVM's extensible full-stack framework enables deep learning applications to efficiently deploy across an array of hardware modules, platforms, and systems, including mobile phones, wearables, specialized chips, and embedded devices. Features include:

High Performance: compilation and minimal runtimes commonly unlock ML workloads on existing hardware.

Runs Everywhere: automatically generates and optimizes tensor operators on backends, CPUs, GPUs, browsers, microcontrollers, FPGAs, ASICs, and more.

Flexible: deep learning compilation models in Keras, Apache MXNet (incubating), PyTorch, Tensorflow, CoreML, and DarkNet, among other libraries. Supports block sparsity, quantization, random forests/classical ML, memory planning, MISRA-C compatibility, Python prototyping, and more.

Easy to Use: easily build out production stacks using C++, Rust, Java, or Python. Deploy deep learning workloads across diverse hardware devices.

Apache TVM is in use at dozens of organizations and institutions that include Alibaba Cloud, AMD, ARM, AWS, Carnegie Mellon University, Cornell University, Edge Cortix, Facebook, Huawei, Intel, ITRI, Microsoft, NVIDIA, Oasis Labs, OctoML, Qualcomm, University of California/Berkeley, UCLA, University of Washington, Xilinx, and more.

"ML compilers and runtimes thrive on diversity of models supported and HW targets, which is a perfect way to show the power of Open Source communities," said Luis Ceze, CEO of OctoML and Professor at the University of Washington. "It has been fantastic to see Apache TVM’s fast adoption among hardware vendors and ML end-users, being well on its way to becoming a de-facto industry standard."

"Apache TVM brings unique value to deep learning researchers and developers. It closes the gap between model development and the demand to efficiently deploy it on various hardware targets," said Yizhi Liu, Senior Software Development Engineer at AWS and member of the Apache TVM Project Management Committee. "I'm thrilled to see Apache TVM now becomes the Top-Level Project and looking forward to further collaboration with the community."

"Congratulations to the Apache TVM community for graduating to be one of the Top Level Projects of The Apache Software Foundation," said Henry Saputra, ASF Member and Apache TVM Incubating Mentor. "The Apache TVM ecosystem has a healthy mix of representation and contribution from the industries and academia that provides a good balance of innovations and production readiness for wider and faster adoption. As one of the mentors of the podling, I am grateful and glad to be part of the journey."

"The key to Apache TVM's success is its open community," added Chen. "We welcome everyone interested in the field to join us and shape the future of ML compilation together under The Apache Way."

Catch Apache TVM in action at the annual TVM Conference being held 2-4 December 2020. The online event is free of charge to participate: for more information and to register, visit https://tvmconf.org/

Availability and Oversight