AI Paper Atlas

Research map / Architectures and efficiency / Quantization

QuIP: 2-Bit Quantization of Large Language Models With Guarantees

Jerry Chee, Yaohui Cai, Volodymyr Kuleshov, Christopher De Sa

NeurIPS 2023 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add incoherence processing (LDLQ to QuIP) at 2 bits (ablation)LDLQ/OPTQ · WikiText2 OPT-30bperplexity: 71.70 -> 11.48Improved
add incoherence processing (LDLQ-RG to QuIP-RG) at 2 bits (ablation)LDLQ-RG · WikiText2 OPT-30bperplexity: 49.40 -> 11.68Improved
add incoherence processing to greedy updates at 2 bits (ablation)Greedy · WikiText2 OPT-30bperplexity: 4,817 -> 11.59Improved
add incoherence processing to nearest rounding at 2 bits (ablation)Near · WikiText2 OPT-30bperplexity: 41,548 -> 12.04Improved
add incoherence processing at 2 bits (ablation)OPTQ · WikiText2 OPT-125mperplexity: 4,563 -> 347.4Improved
add incoherence processing at 3 bits (ablation)OPTQ · WikiText2 OPT-6.7bperplexity: 14.87 -> 11.51Improved
add incoherence processing at 3 bits (ablation)OPTQ · WikiText2 OPT-2.7bperplexity: 17.09 -> 17.44Hurt
add incoherence processing at 4 bits (ablation)OPTQ · WikiText2 OPT-125mperplexity: 31.44 -> 33.35Hurt
add incoherence processing at 4 bits (ablation)OPTQ · WikiText2 OPT-2.7bperplexity: 12.93 -> 12.39Improved
replace nearest rounding with unbiased stochastic rounding in LDLQ/OPTQ (ablation)LDLQ/OPTQ · OPT 125m, baseline processing 2-bitperplexity difference: 241.3Hurt
use greedy updates alone instead of LDLQ in proxy loss at 2 bits (ablation)Greedy · proxy lossproxy loss: 2,554.89 -> 2,587.17No clear effect
use nearest rounding instead of LDLQ in proxy loss at 2 bits (ablation)Near · proxy lossproxy loss: 2,554.89 -> 5,971.69Hurt

Later papers that compared with this method

19 later accepted papers report QuIP as a baseline in their results tables; 19 of them report a better number on most of the results they share with it.

11 more not listed.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy