Event
M.S. Thesis Defense: Rayyan Abdalla
Thursday, August 13, 2026
1:00 p.m.
AVW 2460
Emily Irwin
301 405 0680
eirwin@umd.edu
ANNOUNCEMENT: MS Thesis Defense
Name: Rayyan Abdalla
Committee:
Professor Min Wu, Chair/Advisor
Professor Dinesh Manocha, Co-Advisor
Professor Ramani Duraiswami
Date/time: Thursday, August 13 at 1:00 PM
Location: AVW 2460
Zoom Link: https://umd.zoom.us/j/
Title: Adaptive Post-Training Quantization for Language-centric Multimodal Foundation Models
Abstract:
The growing scale and complexity of language-centric foundation models spanning text, vision, and speech impose substantial storage, memory, and inference costs, motivating post-training quantization (PTQ) for efficient deployment. However, architectural and modality-specific differences govern how quantization affects internal representations and downstream performance, motivating complementary quantization approaches for distinct model structures and sensitivity profiles.
This thesis develops separate quantization strategies for weight-centric and joint weight–activation-sensitive architectures. Saliency-Aware Graph-guided Efficient Post-Training Quantization (SAGE-PTQ) targets models where weight compression is the primary objective. SAGE-PTQ preserves salient weights at higher precision while binarizing graph-guided groups of unsalient weights, enabling near-binary representations with low auxiliary overhead. In contrast, Quantization via Adaptive Migration (QAM) targets speech-to-text architectures that are particularly sensitive to both weight and activation characteristics. QAM jointly leverages weight and activation statistics to identify sensitive input channels and adaptively redistribute quantization resolution while maintaining uniform low-bit weights.
Across large language, vision-language, and speech-to-text models, the proposed methods preserve task performance while substantially reducing model storage and device-memory requirements. These results demonstrate the importance of adapting quantization strategies to the architectural characteristics and modality-specific sensitivity profiles of language-centric multimodal models.
