GGBond stands for GGML Bounding, is a simple and naive GGML Python binding via pybind11. Working in progress.
pip install -e .- Python 3.10+
- CMake 3.15+
- C++17 compatible compiler
GGBond provides three layers of abstraction:
A near 1:1 mapping of the GGML C API. Functions follow GGML naming conventions (new_tensor_2d, mul_mat, backend_graph_compute, etc.). All GGML opaque pointers are wrapped as distinct Python types (ggml.Context, ggml.Tensor, ggml.Backend, etc.) for type safety. Use this layer when you need full control over the GGML computation model.
Object-oriented wrappers around GGML primitives with lifecycle management (close() / context manager). They simplify common patterns but are not a complete 1:1 equivalent of the raw API — for example, Graph internally owns its own Context for graph operations, and Context computes memory size from n_tensors automatically. These are primarily used as building blocks for the higher-level API.
Session owns a backend and manages all resource lifetimes. Tensor is a lazy-evaluated tensor bound to a session — operations build a computation graph, which is materialized on compute() or numpy(). GGUF model weights are loaded directly onto the target backend (CPU/Metal) without intermediate copies.
import numpy as np
import ggbond
matrix_a = np.array([[2, 8], [5, 1], [4, 2], [8, 6]], dtype=np.float32)
matrix_b = np.array([[10, 5], [9, 9], [5, 4]], dtype=np.float32)
s = ggbond.Session("cpu")
a = s.tensor(matrix_a)
b = s.tensor(matrix_b)
print((a @ b).numpy())
s.close()examples/simple.py— Matrix multiplicationexamples/magika.py— File type detection with Magikaexamples/gpt2.py— GPT-2 text generation
MIT