What is TVM
TVM is an evolving open Deep Learning compiler stack. It aims to close the gap between Deep Learning frameworks like TensorFlow, PyTorch, and hardware backends including CPU, GPU, etc. In order to do that, TVM introduces an Intermediate Representation called RelayIR. TVM will first convert Deep Learning model to RelayIR, and then apply optimizations based on RelayIR, finally compile it into executable binaries for target backend, such as LLVM and CUDA. TVM already supports Deep Learning models from many popular frameworks like TensorFlow, ONNX, etc.
Just like TensorFlow and other Deep Learning frameworks, TVM provides user-friendly python APIs and efficient C++ based kernel.
TVM applies lots of optimization for Deep Learning inference on both graph level and operator level. We will introduce them in following sections.
References
[1] Tianqi Chen, et al. TVM: An automated end-to-end optimizing compiler for deep learning. In 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), Carlsbad, CA, 2018. USENIX Association.
Last updated