لطفا منتظر بمانید ...
0% Complete
صفحه اصلی
/
سی و چهارمین کنفرانس بین المللی مهندسی برق
Efficient Large Language Models with Zero-Shot Adjustable Acceleration
نویسندگان :
Sajjad Kachuee
1
Mohammad Sharifkhani
2
1- دانشگاه صنعتی شریف
2- دانشگاه صنعتی شریف
کلمات کلیدی :
Zero-Shot Adjustable Acceleration،Preservation Rate،Hidden Activation،Context Contribution
چکیده :
Large Language Models (LLMs) face significant challenges in real-world applications, particularly in balancing computational efficiency and model performance. Acceleration during both fine-tuning and inference is essential for designing efficient architectures. This paper presents a novel training and inference method, termed zero-shot adjustable acceleration, which dynamically adjusts hardware utilization at inference time without additional fine-tuning. The method is applied to recent LLMs and evaluated on multiple classification and text generation benchmarks. Experimental results demonstrate that the approach enables flexible zero-shot acceleration and achieves up to an 11× speedup over baseline methods, while maintaining high model performance.
لیست مقالات
لیست مقالات بایگانی شده
CEU-Net: A Connected EfficientUNet with ASPP for Mammographic Lesion Segmentation
Fatemeh Jafari - Saeed Meshgini - Reza Afrouzian
Multi-physics electromagnetic-mechanical analysis of a high-speed switched reluctance motor for vacuum cleaner application
Nasrin Majlesi - Morteza Saghaian-Nejad - Amir Rashidi
Low-cost Broadband Reflectarray Antenna Using Cross Bow-Tie elements
Mahdieh Bozorgi - Mahmood Rafaei-Booket
An Energy Compensation Approach to Variable Stiffness Single Leg Jumping
Mohammadreza Gilak - Mehdi Zarei - Mohammad Hossein Basiri - Saman Samiei - Mohammad Mehdi Jalalmaab - Behzad Ahi
A Compact, Highly-Selective, Smart Frequency Detectection Method Using Acoustic-Wave Lumped-Element Resonator Filter
Alireza Azar - Mohammad Khoshakhlagh - Nasser Masoumi
Gearbox Fault Detection Using Continuous Wavelet Transform and Vision Transformer (ViT)
Ali Asadian - Yassin Riyazi - Moosa Ayati
Dual-Arm Tabletop Rearrangement with Movable Stacking and Gap-Aware Push Transfers
Mohammadreza Firoozalizadeh - Zakiye Rostamirad - Arman Barghi - Ahmad Kalhor - Mohammadreza Nayeri - Mehdi Tale Masouleh
Peer-to-peer Energy Sharing Considering Prosumers' Preferences and Load Uncertainties
Mohammad Bagher Moradi - Mohammad Hasan Nazari - Seyed Hossein Hosseinian - Hamed Nafisi
Switched-Inductor Cuk and SEPIC Power Factor Correction Rectifiers
Maryam Pourmahdi-torghabe - Hamed Heydari-doostabad - Reza Ghazi
A High Gain, High IIP3, Perfect Input Matching, Programmable Gain LNA in CMOS Technology
Amirhossein Tajik - Seyedali Samareh Taherinasab - Samad Sheikhaei
بیشتر
ثمین همایش، سامانه مدیریت کنفرانس ها و جشنواره ها - نگارش 44.7.2
