Stacks Deep Learning Compiler

Early release of stacks deep learning compiler based on TVM for System Stacks

Tensor Virtual Machine or TVM compiler image is a layered deep learning graph compiler that is optimized for CPUs and have runtime bindings for Python.

Front end libraries can interface with the compiler, loading pretrained models, to be compiled and deployed using TVM. There are 2 levels of optimizing phases in TVM, the first one is when a model is loaded, where graph level optimizations such as layer fusion and layout transformation are attempted. The next phase includes operator level optimization and code generation including a specialized operator generation using an intelligent scheduler, please refer to the paper for more details.


To install front-end deep learning libraries, use:


This will install Pytorch, TorchVision and ONNX.

Smoke tests

Run some core unit and functional tests:

cd ./tvm/tests
docker run -it -v`pwd`:/workspace dlrs-ml-compiler