TRIAD-PTQ is a post-training quantization scheme designed specifically for edge-class neural networks (under 300M to approximately 3B parameters): small CNNs such as MobileNetV2/V3 and EfficientNet-Lite, compact language models such as TinyLlama-1. 1B, Llama-3. 2-1B/3B, Phi-2, Qwen2. 5-0. 5B/1. 5B, and SmolLM, and small vision-language models such as PaliGemma, MobileVLM, TinyLLaVA, and SmolVLM. The method combines three components: a global Hessian sensitivity router built on KFAC factorization with an empirical inter-layer propagation coefficient, a data-free super-weight identifier preserved at FP16 (top 0. 01 to 0. 1 percent of weights), and an analytic activation-weight cross-covariance grid that generalizes the per-channel scaling of AWQ and SmoothQuant through eigendecomposition of the empirical activation covariance with a closed-form smoothing exponent. A single call (model = triadₚtq (model, bits=4) ) runs in minutes on a consumer GPU, requires 64 to 512 calibration samples, and performs no backward passes. Quantization is weight-only (W4 or W3) ; the method does not provide lossless compression. Expected degradation is bounded analytically through a KFAC-based functional error theorem with explicit falsification criteria. Hardware execution targets NVIDIA GPUs as the primary platform (H100, A100, RTX 4090, Jetson Orin), with explicit deployment paths for Samsung Mali GPU and Exynos NPU (Galaxy SoCs from Exynos 9820 to 2400/2500), Apple Silicon via MLX, and Qualcomm Hexagon via QNN. This is a theoretical proposal with experimental validation protocol. All numerical results in the document are clearly labeled as projections derived from extrapolation of cited prior work, not measurements.
Artem Katolikov (Sun,) studied this question.