Follow the instruction of the original TVM. It should be built with CUDA and LLVM support. To do that, modify config.cmake to change set(USE_CUDA OFF) to set(USE_CUDA ON) and set(USE_LLVM OFF) to set(USE_LLVM ON).
Also, add the following to config.cmake to enable LLIS support:
set(USE_LLIS /path/to/llis)
Inside /path/to/llis, it should include a directory called include which contains the headers of LLIS, and a directory called lib, which contains the libraries of LLIS.
After installing, manually copy some header files that CMake fails to install:
cp -r 3rdparty/dmlc-core/include/dmlc "${INSTALL_PREFIX}/include"
cp -r 3rdparty/dlpack/include/dlpack "${INSTALL_PREFIX}/include"
When executing a TVM script that targets cuda-llis, the environment variable LLIS_PATH should be set to /path/to/llis.