.venv/transformers/docs/source/en/quantization/spqr.md

SpQR

SpQR quantization algorithm involves a 16x16 tiled bi-level group 3-bit quantization structure, with sparse outliers as detailed in SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression.

To SpQR-quantize a model, refer to the Vahe1994/SpQR repository.

Load a pre-SpQR-quantized model in [~PreTrainedModel.from_pretrained].

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

quantized_model = AutoModelForCausalLM.from_pretrained(
    "elvircrn/Llama-2-7b-SPQR-3Bit-16x16-red_pajama-hf",
    torch_dtype=torch.half,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("elvircrn/Llama-2-7b-SPQR-3Bit-16x16-red_pajama-hf")