DeltaZip: Multi-Tenant Language Model Serving via Delta Compression
–arXiv.org Artificial Intelligence
Fine-tuning large language models (LLMs) for downstream tasks can greatly improve model quality, however serving many different fine-tuned LLMs concurrently for users in multi-tenant environments is challenging. Dedicating GPU memory for each model is prohibitively expensive and naively swapping large model weights in and out of GPU memory is slow. Our key insight is that fine-tuned models can be quickly swapped in and out of GPU memory by extracting and compressing the delta between each model and its pre-trained base model. We propose DeltaZip, an LLM serving system that efficiently serves multiple full-parameter fine-tuned models concurrently by aggressively compressing model deltas by a factor of $6\times$ to $8\times$ while maintaining high model quality. DeltaZip increases serving throughput by $1.5\times$ to $3\times$ and improves SLO attainment compared to a vanilla HuggingFace serving system.
arXiv.org Artificial Intelligence
Dec-8-2023
- Country:
- Oceania > Australia (0.04)
- South America
- North America
- Canada (0.04)
- United States
- South Carolina (0.04)
- District of Columbia > Washington (0.04)
- North Carolina
- Craven County > New Bern (0.04)
- Wake County > Raleigh (0.04)
- New York > New York County
- New York City (0.04)
- California
- Alameda County > Berkeley (0.04)
- San Diego County > San Diego (0.04)
- Europe
- United Kingdom (0.45)
- Italy (0.04)
- Russia (0.04)
- Germany (0.04)
- Switzerland > Zürich
- Zürich (0.04)
- Asia
- Nepal (0.04)
- Russia (0.04)
- India (0.04)
- China > Tibet Autonomous Region (0.04)
- Myanmar > Tanintharyi Region
- Dawei (0.04)
- Middle East
- Iran (0.28)
- Jordan (0.04)
- Republic of Türkiye > Batman Province
- Batman (0.04)
- Genre:
- Personal > Interview (1.00)
- Research Report > New Finding (0.67)
- Industry:
- Automobiles & Trucks > Manufacturer (1.00)
- Information Technology > Services (1.00)
- Food & Agriculture > Agriculture (1.00)
- Leisure & Entertainment (1.00)
- Energy > Renewable (1.00)
- Banking & Finance > Trading (1.00)
- Consumer Products & Services (0.92)
- Law > Civil Rights & Constitutional Law (0.67)
- Education > Educational Setting (0.67)
- Health & Medicine
- Consumer Health (1.00)
- Pharmaceuticals & Biotechnology (0.67)
- Health Care Providers & Services (0.67)
- Therapeutic Area
- Neurology (1.00)
- Psychiatry/Psychology > Mental Health (0.92)
- Cardiology/Vascular Diseases (0.92)
- Infections and Infectious Diseases (0.67)
- Government
- Post Office (0.67)
- Foreign Policy (0.67)
- Regional Government > North America Government
- United States Government (1.00)
- Transportation
- Infrastructure & Services (1.00)
- Ground > Road (1.00)
- Passenger (0.67)
- Media
- Television (1.00)
- Film (0.67)
- Technology: