pare-quant 0.1.3

Production-ready quantization for large language and multimodal models

Pare

Quantize any LLM in one line. Switch between GPTQ, AWQ, SmoothQuant, and RTN by changing a config field.

---

Benchmarks

WikiText-2 perplexity (PPL ↓), A40 46 GB:

Method	Llama-3.1-8B	Qwen2.5-7B	OLMo-3-7B
FP16 baseline	6.24	6.85	9.92
RTN INT8	6.25 (+0.01)	6.85 (+0.00)	9.92 (+0.00)
GPTQ INT4	11.10 (+4.86) ‡	7.04 (+0.19)	10.21 (+0.29)
AWQ INT4	6.77 (+0.53)	7.13 (+0.28)	10.36 (+0.44)

Zero-shot accuracy — 6-task average (LAMBADA, PIQA, WinoGrande, OpenBookQA, RTE, COPA) ↑:

Method	Llama-3.1-8B	Qwen2.5-7B	OLMo-3-7B
FP16 baseline	73.22	74.13	69.57
RTN INT8	73.00 (−0.22)	74.09 (−0.04)	69.57 (0.00)
GPTQ INT4	71.65 (−1.57)	73.39 (−0.74)	69.42 (−0.15)
AWQ INT4	70.69 (−2.53)	73.93 (−0.20)	69.57 (0.00)

Throughput at BS=1 (tok/s), dequantize-on-the-fly †:

Method	Llama-3.1-8B	Qwen2.5-7B	OLMo-3-7B
FP16	25.8	32.4	25.1
RTN INT8	2.1	2.3	2.3
GPTQ INT4	1.1	1.2	1.2
AWQ INT4	1.1	1.2	1.2

† With the optional Triton kernel: 8.8× faster at BS=1, 2.8× at BS=4.

‡ Llama-3.1-8B is sensitive to column ordering. With act_order=True: PPL improves from 11.10 to 6.54 (+0.30), accuracy moves from 71.65 to 70.05. Qwen2.5-7B and OLMo-3-7B are unaffected (PPL: 7.04 to 7.02, 10.21 to 10.16).

---

Installation

pip install pare-quant                   # latest
pip install pare-quant==0.1.0           # pin to specific version
pip install "pare-quant[all]"            # + transformers, datasets, Triton kernel

Python ≥ 3.11 · PyTorch ≥ 2.1

---

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
from pare import quantize, QuantConfig
from pare.calibration.data import load_wikitext2_calibration

model     = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B")
calib     = load_wikitext2_calibration(tokenizer, n_samples=128, seq_len=2048)
# or use your own: a list of tokenized tensors of shape (seq_len,)

# Default is AWQ. Change scheme= to switch methods.
config = QuantConfig(bits=4, scheme="awq", group_size=128)   # ← swap to "gptq", "rtn", "smoothquant"
model  = quantize(model, config, calibration_data=calib, device="cuda")

Save and reload:

from pare import save_quantized, load_quantized

save_quantized(model, "qwen25-awq-int4/")
# [pare] Saved 224 quantized layers to qwen25-awq-int4  (3821 MB)

from transformers import AutoConfig, AutoModelForCausalLM
config = AutoConfig.from_pretrained("Qwen/Qwen2.5-7B")
model  = AutoModelForCausalLM.from_config(config)   # architecture only, no weights
model  = load_quantized(model, "qwen25-awq-int4/")

---

Methods

scheme=	Calibration	Quality	When to use
"awq" ★	Yes	★★★★	Default. Best robustness across architectures; recommended starting point
"gptq"	Yes	★★★★	Matches AWQ on Qwen2.5-7B; architecture-agnostic — works correctly across pre- and post-norm models
"smoothquant"	Yes	★★★★	INT8 W+A; closest to FP16 PPL; no INT4
"rtn"	No	★★★	No calibration needed; good baseline or for NF4/FP8

★ Default: QuantConfig() uses AWQ. AWQ is the strongest INT4 method on Qwen2.5-7B (−0.20 vs FP16 baseline). GPTQ is architecture-agnostic and is recommended when the target model's architecture is uncertain. On Llama-3.x architectures, act_order=True is recommended; it reduces PPL from 11.10 to 6.54 on Llama-3.1-8B. On Qwen2.5 and OLMo-3 the effect is negligible.

All schemes support bits=4 or bits=8. Use group_size=128 (default) for best INT4 quality.

Additional options

act_order=True — Sort quantization by activation magnitude (improves GPTQ quality on modern architectures):

QuantConfig(bits=4, scheme="gptq", group_size=128, act_order=True)

Mixed-precision — Automatically promote sensitive layers to higher bits:

QuantConfig(bits=4, scheme="awq", sensitive_bits=8, sensitivity_threshold=0.05)
# [pare] 12 of 224 layers promoted to INT8 based on activation-weighted error

NF4 — Normal float 4-bit codebook (QLoRA-compatible base model format):

from pare.core.dtype import QuantDtype
QuantConfig(bits=4, dtype=QuantDtype.NF4, scheme="rtn")

FP8 — 8-bit float for A100/H100:

QuantConfig(bits=8, dtype=QuantDtype.FP8_E4M3, scheme="rtn")

---

Inference speedup (Triton kernel)

The optional Triton INT4 kernel fuses dequantization into the matmul, avoiding materialising the full FP16 weight matrix. Applies to INT4 schemes (AWQ, GPTQ, RTN). Enable per-layer after quantization:

from pare.layers.linear import QuantizedLinear

for m in model.modules():
    if isinstance(m, QuantizedLinear):
        m.use_kernel = True

Batch size	Without kernel	With kernel	Speedup
1 (decode)	2.09 ms/layer	0.24 ms/layer	8.8×
4	2.18 ms/layer	0.78 ms/layer	2.8×
16	2.66 ms/layer	3.18 ms/layer	0.8×

Requires pip install triton>=3.0.

---

Hardware

	Minimum
Quantizing a 7B model	20 GB VRAM (layerwise strategy peaks at ~2 GB)
RTN / GPTQ / AWQ / NF4	Any CUDA GPU
SmoothQuant W+A	Any CUDA GPU
FP8	PyTorch ≥ 2.1 (A100 via software; H100 native)
Triton kernel	CUDA GPU + triton ≥ 3.0

---

References

The methods implemented in Pare are from the following papers:

• GPTQ — Frantar, Ashkboos, Hoefler, Alistarh. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers. ICLR 2023. arXiv:2210.17323
• AWQ — Lin et al. AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration. MLSys 2024. arXiv:2306.00978
• SmoothQuant — Xiao, Lin, Seznec, Wu, Demouth, Han. SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models. ICML 2023. arXiv:2211.10438
• KIVI — Liu, Yuan, Jin, Zhong, Xu, Braverman, Chen, Hu. KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache. ICML 2024. arXiv:2402.02750
• NF4 / QLoRA — Dettmers, Pagnoni, Holtzman, Zettlemoyer. QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023. arXiv:2305.14314

Citation

@misc{moslem2026pare,
  author = {Moslem, Yasmin},
  title  = {Pare: Production-ready quantization for large language models},
  year   = {2026},
  url    = {https://github.com/TinyAdapt/Pare},
}