Installation

This page covers the PyPI quick install, local build prerequisites, CMake configuration, optional Python support, and the docs site.

Quick Install (PyPI)

For Python users on a machine with CUDA already set up:

pip install loopformer

This builds the Python bindings and CLI from source via scikit-build-core. Requirements:

• nvcc on PATH (CUDA 12+ recommended)
• an NVIDIA GPU (default target sm_90 for H100)
• Python 3.9+

To target a different GPU architecture:

CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=80" pip install loopformer

If you want the full C++ framework library, CLI binary, and test targets, follow the source-build steps below instead.

Build Prerequisites

TinyLoop requires:

• CMake 3.18 or newer
• a CUDA toolkit installation with nvcc
• a C++17-capable host compiler
• Python 3 for tooling and optional bindings

Optional components:

• pybind11 for tinyloop_py
• Node.js for the Docusaurus docs site

CUDA Architecture Selection

TinyLoop defaults to CMAKE_CUDA_ARCHITECTURES=90, which targets H100-class hardware.

Common values:

GPU family	Architecture
A100	80
H100	90

If you do not set this explicitly, the repo defaults to H100.

Configure

From the repository root:

cmake -S tinyloop -B tinyloop/build -DCMAKE_CUDA_ARCHITECTURES=90

If pybind11 is installed and CMake can find Python3 Development.Module, the Python binding will also be configured automatically.

Build

cmake --build tinyloop/build -j

This produces:

• tinyloop/build/tinyloop
• tinyloop/build/libtinyloop_framework.a
• tinyloop/build/tinyloop_py.* when Python binding support is available
• test binaries under tinyloop/build/

Install

cmake --install tinyloop/build --prefix /tmp/tinyloop-install

Installed contents:

• the TinyLoop::framework CMake target
• public headers under include/tinyloop
• package metadata under lib/cmake/TinyLoop
• the tinyloop CLI binary

Using TinyLoop From Another CMake Project

find_package(TinyLoop REQUIRED)

add_executable(my_runtime main.cpp)
target_link_libraries(my_runtime PRIVATE TinyLoop::framework)

Python Binding Build Notes

The binding is built only when both conditions are true:

• pybind11 is found by CMake
• Python3 Development.Module is available

When that happens, the module surface includes:

• Model.score()
• Model.score_last()
• Model.generate()
• Model.generate_speculative()
• prefix-cache helpers

Documentation Site

The docs site lives under tinyloop/wiki.

Install dependencies:

cd tinyloop/wiki
npm ci

Run locally:

npm run start

Build statically:

npm run build

Recommended First Validation

After the build:

1. run inspect on a real .tinyloop artifact
2. run a small benchmark
3. run ctest --test-dir tinyloop/build --output-on-failure

If you have a real model path available:

TINYLOOP_TEST_MODEL_PATH=/absolute/path/to/model.tinyloop \
ctest --test-dir tinyloop/build --output-on-failure

Common Setup Failures

Failed to find nvcc

The active shell cannot see the CUDA toolkit. Check:

• nvcc is installed
• CUDA binaries are on PATH
• CUDAToolkit_ROOT points at the intended toolkit

Python module did not build

Check:

• pybind11 is installed in a CMake-visible location
• Python dev headers are present
• CMake output reports Python3_Development.Module_FOUND