PrismQuant-Llama-3.1-70B

Paper · Code · Loading guide

Official checkpoints for PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers. PrismQuant aligns dominant activation directions with the constant directions of asymmetric quantization groups.

Models

Model key	Base model	Quantization
`llama31_70b`	unsloth/Meta-Llama-3.1-70B	W4A4KV4

Quick start

git clone --depth 1 https://github.com/ForeverBlue816/PrismQuant.git
cd PrismQuant
pip install -e .

from prismquant import load_model

loaded = load_model("llama31_70b", device_map="balanced")
print(loaded.generate("The key idea behind quantization is", max_new_tokens=64))

The loader selects the default PrismQuant checkpoint, downloads its required files at a pinned revision, and obtains the matching base model and tokenizer. These are base models for text completion.

Checkpoint format

The GPTQ INT4 values are stored as dequantized floating-point tensors. The reference runtime applies activation and KV quantization while retaining floating-point storage. Use the PrismQuant loader; this repository is not a standalone Transformers checkpoint or a packed INT4 serving model.

This fp32 reference requires roughly 280 GB for model parameters, plus runtime overhead. Use sufficient aggregate GPU memory and, if needed, max_memory={...} with balanced placement.

config.json declares the PrismQuant artifact format and default models. checkpoints/ contains decoder weights; rotations/ contains the factors required by the loader.