Installation

Install the components for your own agent workflow or a kcoral GPU server.

For optimization runs, setup prepares the packages and skills automatically.

To run an exported kernel with tvm-ffi, follow Export a kernel for the consumer dependencies.

Before you start

  • Linux x86_64, Python 3.12 or 3.13, and pip 25.1 or later.

  • The CUDA Toolkit (nvcc, ptxas) where GPU kernels are compiled, and a compatible NVIDIA driver where they are run.

  • Cargo, Rust 1.95.0 or later, and a C linker for numerical simulation and correctness checks with either installation method.

Install Python packages

Activate your agent’s Python environment and choose one method.

Install from PyPI

Install the released package without cloning this repository:

python -m pip install tirx-harness

Wheels include the native frontend. If pip builds from a source distribution, the build tools below are required.

Build from source

Source builds require Git, C/C++ build tools, Python development headers, and the Rust toolchain listed above.

git clone https://github.com/mlc-ai/TIRx-harness.git
cd TIRx-harness
python -m pip install .

Optional: install with uv

From the initialized checkout, build the harness with dependencies from uv.lock:

uv sync --locked
source .venv/bin/activate

Verify the installation

After any of these methods, check imports (this does not run checks or GPU kernels):

python -c "import tvm_tirx, tvm_ffi, tirx_kernels.tirx_lite, tirx_harness; print('Core imports OK')"

Install kcoral server dependencies

From the repository root, run python -m pip install --group server, or uv sync --locked --only-group server and activate .venv. These install only the server dependencies. See remote execution to start the server.

Install agent skills

Skill

Purpose

tirx-wiki

Find TIRx-lite APIs, canonical kernels, and GPU references.

tirx-debug-kernel

Check correctness, investigate findings, and verify fixes.

tirx-profile-kernel

Measure performance and use profiler evidence to guide changes.

The package bundles the skills. Install them into the directory your agent reads, such as .agents/skills or .claude/skills:

tirx-harness skills install --dest /absolute/path/to/your/project/.agents/skills

The command also downloads the wiki manuals and reference repositories, including tirx-kernels. Use --no-fetch to skip the download and --force to replace an earlier installation. Editable installs, such as uv sync, do not bundle the skills; copy them from skills/ in the checkout instead.

Continue to Quick Start for a concrete example.