Llama-3.1-70B-Instruct-2Bit-GSQ

Llama-3.1-70B-Instruct — 2-bit GSQ

2-bit quantization of meta-llama/Llama-3.1-70B-Instruct produced with GSQ (Gumbel-Softmax Quantization) at ≈2.13 bpp.

GSQ is the strongest scalar PTQ method we measured at this scale and lands within ≈1.7 points of vector-quantized methods (QTIP, PV-Tuning) on the standard zero-shot suite (ARC-C/E, HellaSwag, PIQA, Winogrande):

• Paper: GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (arXiv:2604.18556)
• Paper page on HF: https://huggingface.co/papers/2604.18556
• Code: https://github.com/IST-DASLab/GSQ
• Collection: https://huggingface.co/collections/ISTA-DASLab/gsq

Quantization details

• Base model: meta-llama/Llama-3.1-70B-Instruct
• Bits / weight (effective): ≈2.13 bpp
• Codebook: 2-bit symmetric scalar {-2, -1, 0, +1} × scale
• Group size: 128
• Format: Humming (quant_method: "humming", b_dtype: "uint2")
• Pipeline: GPTQ initialization → Gumbel-Softmax refinement (Lion optimizer)

Storage layout (why the HF UI shows I32 + BF16)

The Hugging Face "Tensor types" widget reports the container dtype of each safetensor on disk, not the effective precision of the underlying weights. This checkpoint uses the Humming on-disk layout (exact-width packing — no sub-byte values are padded into a wider container). For every quantized Linear layer with original weight shape [out_features, in_features], the following tensors are stored:

Example (model.layers.0.mlp.gate_proj, original [28672, 8192]): weight = [28672, 512] I32 (since 8192 × 2 / 32 = 512), weight_scale = [28672, 64] BF16 (since 8192 / 128 = 64).

So although the UI says "I32 + BF16", the effective storage per quantized weight is 2 bits (packed) + 16 bits / 128 (group scale) ≈ 2.13 bpp. The quantization_config block in config.json is:

{
  "quant_method": "humming",
  "b_dtype": "uint2",
  "weight_scale_group_size": 128,
  "weight_scale_type": "group",
  "has_zero_point": false,
  "ignore": ["lm_head", "embed_tokens"]
}

Loading this checkpoint requires vLLM plus the humming kernels (pip install humming-kernels). See Serving with vLLM below.

Note: GSQ training first writes shards in compressed-tensors pack-quantized format (where a sub-4-bit codebook is padded into a 4-bit INT32 container). The published checkpoint here has been re-packed via convert_to_humming.py into exact-width 2-bit Humming storage, hence the 2 / 32 shape factor you see above.

Serving with vLLM

Install the Humming kernels (required for vLLM to load this checkpoint):

pip install humming-kernels

vllm serve ISTA-DASLab/Llama-3.1-70B-Instruct-2Bit-GSQ \
  --tensor-parallel-size 2

Citation

@article{gsq2026,
  title  = {GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling},
  author = {Dadgarnia, Alireza and Tabesh, Soroush and Nikdan, Mahdi and Helcig, Michael and Kurti{\'c}, Eldar and Kleinegger, Max and Alistarh, Dan},
  journal= {arXiv preprint arXiv:2604.18556},
  year   = {2026},
  url    = {https://arxiv.org/abs/2604.18556}
}