pip install sinq

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, SinqConfig

model_name = "Qwen/Qwen3-1.7B"

cfg = SinqConfig(
    nbits=4,
    group_size=64,
    tiling_mode="1D",
    method="sinq",
    modules_to_not_convert=["lm_head"]
)

tok = AutoTokenizer.from_pretrained(model_name)
qmodel = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=cfg,
    dtype=torch.bfloat16
)

# Save sinq quantized model
model.save_pretrained("/path/to/save/qwen3-1.7B-sinq-4bit")
model.push_to_hub("HF_Hub_username/qwen3-1.7B-sinq-4bit")
tokenizer.push_to_hub("HF_Hub_username/qwen3-1.7B-sinq-4bit")

from sinq.hf_io import patch_hf_pretrained_io
patch_hf_pretrained_io()
# Reload a sinq quantized model
hf_hub_model = "HF_Hub_username/qwen3-1.7B-sinq-4bit"
tokenizer  = AutoTokenizer.from_pretrained(hf_hub_model)
model = AutoModelForCausalLM.from_pretrained(hf_hub_model)

# --- Save to a folder (sharded safetensors) ---

# 'model' must already be SINQ-quantized
# Locally save
qmodel.save_pretrained("/path/to/save/qwen3-1.7B-sinq-4bit")
# Push to the Hub
qmodel.push_to_hub("HF_Hub_username/qwen3-1.7B-sinq-4bit")
tok.push_to_hub("HF_Hub_username/qwen3-1.7B-sinq-4bit")

# --- Reload later--

save_dir = "/path/to/save/qwen3-1.7B-sinq-4bit"
hf_hub_model = "HF_Hub_username/qwen3-1.7B-sinq-4bit"

# From local directory
tok = AutoTokenizer.from_pretrained(save_dir)
qmodel = AutoModelForCausalLM.from_pretrained(save_dir)

# From HF Hub
tok = AutoTokenizer.from_pretrained(hf_hub_model)
qmodel = AutoModelForCausalLM.from_pretrained(hf_hub_model)

import torch
from lm_eval import evaluator
from lm_eval.models.huggingface import HFLM

device = torch.accelerator.current_accelerator().type if torch.accelerator.is_available() else "cpu"

# Wrap the already quantized model and tokenizer with HFLM
lm = HFLM(pretrained=qmodel, tokenizer=tok, device=device)

# Evaluate (many tasks available on lm-eval such as MMLU and HellaSwag)
results = evaluator.simple_evaluate(
    model=lm,
    tasks=["wikitext"],  # small and fast benchmark
    device=device
)

@misc{muller2025sinq,
      title={SINQ: Sinkhorn-Normalized Quantization for Calibration-Free Low-Precision LLM Weights}, 
      author={Lorenz K. Muller and Philippe Bich and Jiawei Zhuang and Ahmet Celik and Luca Benfenati and Lukas Cavigelli},
      year={2025},
      eprint={2509.22944},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={http://arxiv.org/abs/2509.22944}
}