01
Model Compression
Quantization, sparsity, and low-rank decomposition for reducing model redundancy in MLLM inference.
ECCV 2026 Tutorial | Sep 8 AM
Efficient MLLM Inference via Approximated and Exact Computing
A systematic tutorial on efficient MLLM inference.
Overview
Multimodal large language models have advanced rapidly, but their inference cost remains a major barrier for cloud serving and edge deployment. The cost comes from massive model parameters, long multimodal contexts, attention complexity, and memory-bound execution on modern hardware. This tutorial frames efficient MLLM inference through two complementary lenses: approximated computing, which reduces model and data redundancy while preserving practical utility, and exact computing, which accelerates inference through system and hardware optimization without changing model outputs.
Organizers
Organizers
Huan Wang
Westlake University
Can Qin
Ming-Hsuan Yang
UC Merced
Pavlo Molchanov
NVIDIA
Wonmin Byeon
NVIDIA
Vikas Singh
University of Wisconsin-Madison
Li Zhang
Mu Cai
Thinking Machines Lab
Haoxuan You
Meta FAIR
Yang Sui
Microsoft AI
Yuzhang Shang
University of Central Florida
Keda Tao
Westlake University
Kele Shao
Westlake University
Technical Program
Three Main Topic
02
Token Efficiency
Training & training-frree compression methods for long multimodal inputs and autoregressive generation.
03
System-Level Design
Exact computing methods and inference framework that improve inference throughput and latency without changing the computed result.
Speakers
Schedule
Opening Remarks
Huan WangApproximate Computing I: Model Compression
Huan WangApproximate Computing II: Token Efficiency
Bo LiCoffee Break and Discussions
Informal discussion with organizers and speakers.
Exact Computing: System-Level Optimization
Chenyang ZhaoQA and Closing Remarks
Huan WangMaterials
Slides, code, and reading list will be released here.
TBD
Slides
Speaker slides will be posted after the final program is confirmed.
TBD
GitHub Repository
Recommended papers, code links, and reproducible resources.
TBD
Reading List
Curated survey and recent work on efficient MLLM inference.
Affiliations
