HQQ — квантование LLM без калибровочных данных

★ 7.7 · ai-tooling

hqq is a Claude Code skill that implements Half-Quadratic Quantization (HQQ) for LLMs — fast, calibration-free weight quantization supporting 8/4/3/2/1-bit precision across multiple optimized backends. Unlike GPTQ or AWQ, it requires no calibration dataset, so any model can be quantized in minutes rather than hours. Supported backends include PyTorch, pytorch_compile, ATEN, TorchAO int4, Marlin, BitBlas, and GemLite, with native integration into HuggingFace Transformers via HqqConfig and full vLLM compatibility. Quantized models remain fine-tunable through LoRA/PEFT. The skill targets engineers who need to reduce LLM memory footprint quickly or experiment with aggressive compression settings such as 2-bit or 1-bit quantization without access to representative calibration data.