PrimitiveML — Minimal tensor runtime & inference framework in C

PrimitiveML is a minimalist tensor library and inference framework written in C, inspired by PyTorch. It provides a low-level tensor runtime and a modular layer system (Module/Linear/Embedding) with forward()-style calls similar to PyTorch. The repository includes an example of running a small character-based GPT-2 model exported from PyTorch (see `nanogpt).

Features

• Low-level tensor runtime with dynamic shapes, dtypes and strides.
• Tensor ops: reshape, transpose, unsqueeze.
• Element-wise ops with automatic broadcasting, reductions and matrix multiplication.
• Basic activations: ReLU, Sigmoid, Softmax.
• Modular API: Module, Linear, Embedding with forward() semantics.
• Demo: inference of a small char-based GPT-2 model exported from PyTorch.
• CLI for text generation.

Requirements

• C compiler (clang or gcc)
• CMake
• Python + PyTorch (if you want to run example notebook)

Build

From the repository root:

mkdir build && cd build
cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
-DCMAKE_CXX_COMPILER=clang++ \
-DCMAKE_C_COMPILER=clang \
..
make

Project layout

.
├── CMakeLists.txt
├── core/          # tensor implementation
├── include/       # public headers (modules, functional, layers)
├── nanogpt/       # Jupyter notebook and PyTorch model code
├── src/           # C implementation of layers & CLI
└── tests/         # unit tests

Tests

Tests are experimental and incomplete. Some unit tests exist, but coverage is partial and many modules are not yet tested. Use with caution.