Installation¶
Install the components for your own agent workflow or a kcoral GPU server.
For an optimization run, continue to Optimization Runs.
Setup creates the worktree’s .venv and installs its packages and skills
automatically, for both local and remote GPU runs.
Before you start¶
Linux x86_64, Python 3.12 or 3.13, and pip 25.1 or later.
uv for optimization runs.
Git, C/C++ build tools, and Python development headers.
The CUDA Toolkit (
nvcc,ptxas) where GPU kernels are compiled, and a compatible NVIDIA driver where they are run.Cargo and Rust 1.89.0 or later for native checkers.
Get the repository¶
git clone https://github.com/mlc-ai/TIRx-harness.git
cd TIRx-harness
git submodule update --init thirdparty/tvm-rust-ext
Install Python packages¶
For manual installation into your own environment:
Scenario |
Command from the repository root |
|---|---|
Use the harness |
|
Run a kcoral server with benchmark dependencies |
|
After installing the harness, check that its packages resolve:
python -c "import tvm.tirx, tvm_ffi, tirx_kernels, tirx_tools; print('Core imports OK')"
This checks imports only.
Optional: install with uv¶
Use uv to install the versions recorded in uv.lock:
Use the harness:
uv sync --lockedRun a kcoral server with benchmark dependencies:
uv sync --locked --only-group server
Then activate the installed environment:
source .venv/bin/activate
Install agent skills¶
The skills explain which tools to use and how to interpret their results:
Skill |
Purpose |
|---|---|
Find tirx-lite APIs, canonical kernels, and GPU references. |
|
Check correctness, investigate findings, and verify fixes. |
|
Measure performance and use profiler evidence to guide changes. |
For your own agent workflow, copy the skills to the directory your agent reads. From this repository:
skills_dir=/absolute/path/to/your/project/.agents/skills
mkdir -p "$skills_dir"
cp -R skills/tirx-wiki skills/tirx-debug-kernel skills/tirx-profile-kernel "$skills_dir/"
(cd "$skills_dir/tirx-wiki" && python scripts/fetch_references.py)
The fetcher downloads the wiki manuals and reference repositories, including
tirx-kernels with its root README and source docs;
it requires network access. Python dependencies are already installed above.
Use your agent’s discovery convention, such as .agents/skills or
.claude/skills. Copy each skill as a complete directory. For
optimization runs, evolution/setup.py prepares the selected
skills automatically.
Continue to Quick Start for a concrete example.