01
Paper
PaperBitNet: Scaling 1-bit Transformers for Large Language Models
Wang et al. · 2023
Why?
The 1-bit LLM starting point - BitLinear as a drop-in that trains binary weights from scratch.
Quantization
Community reviewedThis shelf has been revised by 1 person so far via rebalance proposals.
For
Engineers who know transformers and want a credible starting set on weight/activation quantization.
Covers
AWQ, SmoothQuant, ZeroQuant, LLM.int8(), QLoRA, LLM-QAT, EfficientQAT, and BitNet / BitNet 1.58.
Leaves out
Hardware kernel design, GPTQ surveys, or non-paper explainers.
The full collection in the curator’s suggested order.
Whole shelf
02
Paper
AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration
Lin et al. · 2024
Why?
Shows why protecting a few salient channels matters more than uniform treatment of every weight.
03
Paper
The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits
Ma et al. · 2024
Why?
BitNet b1.58 - ternary weights that match full-precision quality while rewriting the cost story.
04
Paper
EfficientQAT: Efficient Quantization-Aware Training for Large Language Models
Chen et al. · 2024
Why?
A more practical QAT recipe - block-wise training that keeps accuracy without a full-model retraining bill.
05
Paper
LLM-QAT: Data-Free Quantization Aware Training for Large Language Models
Liu et al. · 2023
Why?
Shows how far you can push QAT for LLMs without a full training corpus - useful contrast to pure PTQ.
06
Paper
LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale
Dettmers et al. · 2022
Why?
The outlier-aware int8 path that made large-model 8-bit inference practical before 4-bit took over.
07
Paper
QLoRA: Efficient Finetuning of Quantized LLMs
Dettmers et al. · 2023
Why?
The paper behind 4-bit finetuning - the method most teams mean when they say QLoRA.
08
Paper
SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models
Xiao et al. · 2023
Why?
Makes activation outliers less catastrophic - essential context once you leave weight-only methods.
09
Paper
ZeroQuant: Efficient and Affordable Post-Training Quantization for Large-Scale Transformers
Yao et al. · 2022
Why?
Early production-minded PTQ stack - layer-wise quantization and affordable calibration at scale.