Megakernel

This repository is a development workspace for TVM/TIR experiments. The root project only contains tooling: the TVM source lives in 3rdparty/tvm, and local scripts build TVM with LLVM and CUDA enabled.

Prerequisites

Install the system dependencies required by TVM's source build:

• Python 3.10+
• Git
• CMake 3.24+
• Ninja
• A C++17 compiler
• LLVM 15+ with llvm-config
• CUDA toolkit with nvcc
• uv

On a fresh checkout, initialize submodules first:

git submodule update --init --recursive

Python Environment

Create or update the uv environment:

uv sync --group dev

Build TVM

The default build follows TVM's official source-build flow: create 3rdparty/tvm/build, copy cmake/config.cmake, append local options, build TVM, install tvm-ffi, then expose 3rdparty/tvm/python to the uv environment with a .pth file. This avoids triggering a second scikit-build CMake build during Python installation.

./scripts/build_tvm.sh

Useful environment variables:

CLEAN=1 ./scripts/build_tvm.sh
JOBS=32 ./scripts/build_tvm.sh
LLVM_CONFIG=/usr/lib/llvm-15/bin/llvm-config ./scripts/build_tvm.sh
CUDA_PATH=/usr/local/cuda ./scripts/build_tvm.sh
CMAKE_BUILD_TYPE=Release ./scripts/build_tvm.sh
CMAKE_GENERATOR="Unix Makefiles" ./scripts/build_tvm.sh

Default TVM options:

• USE_LLVM="<detected llvm-config 15+> --ignore-libllvm --link-static" when all static LLVM libs are present; otherwise the script falls back to --link-shared dynamic linkage.
• USE_CUDA=ON, or USE_CUDA=$CUDA_PATH when CUDA_PATH is set
• USE_CUTLASS=OFF
• USE_CUBLAS=OFF
• USE_CUDNN=OFF
• USE_NCCL=OFF
• USE_NVTX=OFF
• USE_THRUST=OFF
• USE_CURAND=OFF

Validate

After a build, run:

./scripts/check_tvm.sh

The check imports tvm and tvm_ffi, prints the loaded TVM library path, and verifies that the build reports LLVM and CUDA support.

Reference

TVM source build documentation: https://tvm.apache.org/docs/install/from_source.html#install-from-source