Saved in:
Bibliographic Details
Main Authors: Ahmad, Muhammad, Mazher, Khurram, Akram, Saqib, Tameem, Ahmad, Nasir, Saad Bin
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2505.07531
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908534133751808
author Ahmad, Muhammad
Mazher, Khurram
Akram, Saqib
Tameem, Ahmad
Nasir, Saad Bin
author_facet Ahmad, Muhammad
Mazher, Khurram
Akram, Saqib
Tameem, Ahmad
Nasir, Saad Bin
contents We present QuantX: a tailored suite of recipes for LLM and VLM quantization. It is capable of quantizing down to 3-bit resolutions with minimal loss in performance. The quantization strategies in QuantX take into account hardware-specific constraints to achieve efficient dequantization during inference ensuring flexible trade-off between runtime speed, memory requirement and model accuracy. Our results demonstrate that QuantX achieves performance within 6% of the unquantized model for LlaVa-v1.6 quantized down to 3-bits for multiple end user tasks and outperforms recently published state-of-the-art quantization techniques. We further integrate one particular technique from QuantX into the popular Llama.cpp framework and show its feasibility in terms of runtime compared to the mainstream quantization techniques from Llama.cpp. Lastly, this manuscript provides insights into the LLM quantization process that motivated the range of recipes and options that are incorporated in QuantX.
format Preprint
id arxiv_https___arxiv_org_abs_2505_07531
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle QuantX: A Framework for Hardware-Aware Quantization of Generative AI Workloads
Ahmad, Muhammad
Mazher, Khurram
Akram, Saqib
Tameem, Ahmad
Nasir, Saad Bin
Artificial Intelligence
Signal Processing
We present QuantX: a tailored suite of recipes for LLM and VLM quantization. It is capable of quantizing down to 3-bit resolutions with minimal loss in performance. The quantization strategies in QuantX take into account hardware-specific constraints to achieve efficient dequantization during inference ensuring flexible trade-off between runtime speed, memory requirement and model accuracy. Our results demonstrate that QuantX achieves performance within 6% of the unquantized model for LlaVa-v1.6 quantized down to 3-bits for multiple end user tasks and outperforms recently published state-of-the-art quantization techniques. We further integrate one particular technique from QuantX into the popular Llama.cpp framework and show its feasibility in terms of runtime compared to the mainstream quantization techniques from Llama.cpp. Lastly, this manuscript provides insights into the LLM quantization process that motivated the range of recipes and options that are incorporated in QuantX.
title QuantX: A Framework for Hardware-Aware Quantization of Generative AI Workloads
topic Artificial Intelligence
Signal Processing
url https://arxiv.org/abs/2505.07531