9253-GRP Efficient Memory Management for LLM Inference on Constrained Memory Devices
Large Language Model inference on memory-constrained devices suffers from delay due to thrashing. Since models with billions of parameters often cross the memory limit, it becomes imperative to keep the weights on disk and fetch them on demand. As a result, the inference procedure faces delay due to the constant swapping in and out of model parameters between disk and memory. In this paper, we establish this problem by comparing the existing methods in the literature. We also propose a solution to this problem, which is an efficient eviction and prefetching scheme.
Authors
- ARNAB JOY
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00