Is OmniQuant superseded?

OmniQuant (LLM quantization): heavily superseded — a standard baseline that newer methods routinely beat. 2 paper(s) critique it, 14 beat it on benchmarks — #5 of 80 most-superseded. Sub-problem: cluster led by GPTQ. Newer alternatives in the same sub-problem include QVGGT, LFQ, ADMM-Q, OSAQ, SEPTQ.

LLM quantization

OmniQuant

OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models

Heavily superseded#5 of 80 most-superseded · first seen Aug 25, 2023

Heavily superseded — a standard baseline that newer methods routinely beat

2 papers critique it · 14 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites OmniQuant as a baseline.

while AWQ falls apart at even 2.15 bits omniquant and OmniQuant produces unusable models at 2 bits, produces high quality models that are close to OmniQuant 3 bit models.

— QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks

“However, these methods cannot effectively quantize the LLMs to 4-bit weights and activations.”

— Adaptive Layer-Wise Transformations for Post-Training Quantization of Large Language Models

Beaten on benchmarks

Head-to-head results where a newer method reports beating OmniQuant. Values are copied from the source paper's tables — verify against the cited paper.

LieQ beats OmniQuant
71.84 vs 23.75
ARC-e · [Q3-4B 3.00 bits]
Exploring Layer-wise Information Effectiveness for Post-Training Quantization in Small Language Models
Layer-Wise High-Impact Parameter Ratio Optimization beats OmniQuant
14.64 vs 90.64
Perplexity · [LLaMA-2-7B, W2A16]
Layer-Wise High-Impact Parameter Ratio Optimization in Post-Training Quantization for Large Language Models
TesseraQ beats OmniQuant
8.05 vs 37.37
Perplexity · [W2A16]
TesseraQ: Ultra Low-Bit LLM Post-Training Quantization with Block Reconstruction
SEPTQ beats OmniQuant
9.93 vs 30.84
perplexity · [2-bit]
SEPTQ: A Simple and Effective Post-Training Quantization Paradigm for Large Language Models
ButterflyQuant beats OmniQuant
15.40 vs 37.32
WikiText-2 Perplexity · [LLaMA2-7B]
ButterflyQuant: Ultra-low-bit LLM Quantization through Learnable Orthogonal Butterfly Transforms
QuaRot beats OmniQuant
6.10 vs 14.26
PPL · [4-bit, Llama 7B]
QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs
QuIP# beats OmniQuant
6.12 vs 12.28
C4 · [2-70, 2 bits]
QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks
PTQ1.61 beats OmniQuant
37.20 vs 25.35
Avg. · [LLaMA 3-8]
PTQ1.61: Push the Real Limit of Extremely Low-Bit Post-Training Quantization Methods for Large Language Models
SliderQuant beats OmniQuant
8.34 vs 14.26
Wiki · [W4A4]
SliderQuant: Accurate Post-Training Quantization for LLMs
SignRoundV2 beats OmniQuant
57.88 vs 46.98
Avg · [Llama2-7B, 2 bits, Group Size 128]
SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs
ParetoQ beats OmniQuant
61.3 vs 57.1
Avg. · [3-bit quantization on MobileLLM-1B]
ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization
OmniQuant+LFQ beats OmniQuant
79.76 vs 78.17
GSM8K (greedy) · [Llama 3.1 8B, W4, OmniQuant]
LFQ: Logit-aware Final-block Quantization for Boosting the Generation Quality of Low-Bit Quantized LLMs

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.