English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The recent advancements in self-supervised learning, combined with theTransformer architecture, have enabled natural language processing (NLP) toachieve remarkably low perplexity. However, powerful NLP models necessitateincreasing model size, leading to substantial computational and memoryrequirements. In this paper, we introduce an efficient inference frameworktailored for large-scale generative language models. To reduce the model size,we employ a weight-only quantization strategy while preserving full precisionfor activations. As a result, we attain sub-4-bit quantization for each weightthrough non-uniform or uniform quantization techniques. Our proposed kernel,called LUT-GEMM, then accelerates quantized matrix multiplications, offering aflexible balance between compression ratio and accuracy. Unlike earlier matrixmultiplication kernels that accommodated weight-only quantization, LUT-GEMMefficiently eliminates the resource-demanding dequantization process for bothuniform and non-uniform quantization methods. By reducing the latency ofindividual GPUs and the overall inference process for large-scale languagemodels, LUT-GEMM provides significant performance improvements in inference.The impact of LUT-GEMM is facilitated by implementing high compression ratiosthrough low-bit quantization and efficient LUT-based operations, whichdecreases the number of required GPUs. For the OPT-175B model with 3-bitquantization, we show that LUT-GEMM accelerates the latency for generating eachtoken by 2.1x compared to OPTQ, which requires costly dequantization.Consequently, LUT-GEMM enables inference of the OPT-175B model on a single GPUwithout noticeable degradation in accuracy or performance, while thenon-quantized OPT-175B model requires a minimum of 8 GPUs.

LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models