libgguf

libgguf is a standalone GGUF library with native row kernels, Python bindings, Torch and NumPy implementations, optional HIP/ROCm GPU kernels, and a native safetensors-to-GGUF converter.

The native REF CPU backend is the correctness reference. Optimized CPU backends are selected at compile time.

Install

python -m pip install .

Editable development install:

python -m pip install --editable .

For a ROCm build with the Torch GPU wrapper:

$env:CMAKE_ARGS='-DLIBGGUF_BUILD_GPU=ON -DLIBGGUF_ENABLE_ROCM=ON -DLIBGGUF_BUILD_TORCH=ON'
python -m pip install --editable ".[gpu]" --no-build-isolation

See docs/installation.md for build requirements and CMake options.

Python row API

import numpy as np
import libgguf

rows = np.random.default_rng(0).normal(size=(4, 4096)).astype(np.float32)
qtype = libgguf.GGMLQuantizationType.Q4_K

encoded = libgguf.quantize_rows(rows, qtype)
decoded = libgguf.dequantize_rows(encoded, qtype, n_per_row=rows.shape[-1])

The top-level package also exposes qtype metadata, storage-row operations, imatrix loading, lightweight GGUF inspection, raw tensor-byte reads, and structural validation. See docs/python-api.md.

Backends

• libgguf: native CPU row operations and GGUF metadata APIs.
• libgguf.libgguf_numpy: NumPy implementation.
• libgguf.libgguf_torch: Torch-native implementation.
• libgguf.libgguf_gpu: native Torch GPU operations.

GPU builds explicitly select either HIP/ROCm or CUDA. PyTorch exposes ROCm devices through its cuda device namespace:

import torch
import libgguf
import libgguf.libgguf_gpu

rows = torch.randn(4, 4096, device="cuda", dtype=torch.float32)
qtype = libgguf.GGMLQuantizationType.Q4_K
encoded = libgguf.libgguf_gpu.quantize(rows, int(qtype))

Runtime-free device fragments

<libgguf/gpu/dequantize/fragment.cuh> is the lean header-only API for embedding partial block decoding inside downstream GPU kernels. It contains no HIP/CUDA runtime, Torch, launcher, or lookup-table dependency and is also usable from ordinary C++ reference tests. BlockFragmentDecoder<Format> currently exposes scalar and eight-value fragment decoding for Q4_0, Q4_1, Q5_0, Q5_1, and Q8_0. Consumers should pin both the libgguf commit and block_fragment_decoder_api_revision.

Converter and commands

The native libgguf_quantize_gguf executable converts safetensors models to GGUF and supports CPU, GPU, and automatic backend selection.

libgguf_quantize_gguf \
  --src model.safetensors \
  --qtype Q4_K_M \
  --dst model-Q4_K_M.gguf \
  --backend auto

The Python package also installs gguf-inspect, gguf-validate, and gguf-compare. See docs/cli.md.

Documentation

• Qtypes
• Conversion policy
• Correctness
• Documentation index

License

Apache-2.0. Adapted source files retain their applicable provenance notices.