"Guides Apache TVM source/PyPI installation, CMake/LLVM builds,
Scanned 9/9/2026
Install to Claude Code
npx -y skills add VectorSpaceLab/AREX-Skill --skill install-build --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Install Build?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/vectorspacelab-install-build-arex-skill)More formats (shields.io, HTML) on the badges page.
---
name: install-build
description: "Guides Apache TVM source/PyPI installation, CMake/LLVM builds,
import validation, backend probes, and focused test selection."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
---
# Install and Build TVM
Use this route when the task is to install Apache TVM, configure/build a
checkout, validate which libraries Python loads, or choose the smallest native
test after a change.
## Route
1. Decide whether the task needs the PyPI wheel or a source checkout. Use the
source path when changing C++/Python code or enabling custom backends.
2. Read [`references/build-and-test.md`](references/build-and-test.md) for the
build matrix and focused commands.
3. Build with CMake/Ninja and a supported LLVM toolchain. Keep `python/` on
`PYTHONPATH`; do not use editable TVM/tvm-ffi installs for multi-checkout
development.
4. Run [`scripts/check_tvm_install.py`](scripts/check_tvm_install.py) before
broader tests. It reports version, loaded libraries, `libinfo()`, and
backend readiness without changing the checkout.
5. Read [`references/troubleshooting.md`](references/troubleshooting.md) for
stale imports, missing `tvm-ffi`, LLVM/linker failures, or optional extras.
## Minimal source workflow
```bash
# From a TVM checkout with submodules initialized.
git submodule update --init --recursive
mkdir -p build
cmake -S . -B build -G Ninja \
-DCMAKE_BUILD_TYPE=RelWithDebInfo \
-DTVM_BUILD_PYTHON_MODULE=ON \
-DUSE_LLVM="llvm-config --ignore-libllvm --link-static" \
-DUSE_RPC=ON -DUSE_CUDA=OFF
cmake --build build --parallel
export PYTHONPATH="$PWD/python${PYTHONPATH:+:$PYTHONPATH}"
export TVM_LIBRARY_PATH="$PWD/build/lib"
python -c 'import tvm; print(tvm.__version__); print(tvm.support.libinfo())'
```
Adjust `USE_CUDA`, `USE_ROCM`, `USE_VULKAN`, `USE_METAL`, or optional library
flags only after the matching toolkit, headers, driver, and runtime have been
verified. A successful CMake build does not prove a device backend works.
## Boundaries
- Relax module compilation: [`../relax-compile/SKILL.md`](../relax-compile/SKILL.md).
- TIRx authoring and GPU-specific tests: [`../tirx-kernels/SKILL.md`](../tirx-kernels/SKILL.md).
- S-TIR/meta-schedule: [`../s-tir-tuning/SKILL.md`](../s-tir-tuning/SKILL.md).
- RPC service lifecycle and remote execution: [`../rpc-deployment/SKILL.md`](../rpc-deployment/SKILL.md).
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!