Mingjun Xiao

He Sun, Shinan Liu, Li Li et al.

Deploying Large Language Models (LLMs) on memory-constrained AI Personal Computers (AIPCs) enables low-latency, privacy-preserving inference, but long-context generation is fundamentally bottlenecked by the linearly growing Key-Value (KV) cache. While dynamic KV eviction mitigates this memory wall, existing offloading strategies either trigger crippling PCIe I/O bottlenecks on standard SSDs or suffer from FPGA resource exhaustion by forcing compute-intensive exact attention on a single, weak Computational Storage Drive (CSD). In this paper, we propose HillInfer, a CSD-assisted KV eviction framework that introduces a paradigm shift: offloading strictly lightweight token importance evaluation to a single CSD (e.g., SmartSSD) on AIPCs. To fully capitalize on this lightweight offloading strategy, HillInfer orchestrates a Hierarchical KV Cache Manager (HKM) that leverages temporal locality and dynamic token hit rates to physically partition cache pools, thereby eliminating cross-device I/O thrashing. Additionally, we design an Adaptive Prefetch-based Pipeline (APP) that adaptively balances the evaluation workload between the host CPU and the SmartSSD, effectively masking the heterogeneous straggler effect. Finally, we introduce a CSD-based Evaluation Configuration (CEC) to enable resource-efficient near-data processing on the FPGA. Extensive experiments on a commodity AIPC demonstrate that HillInfer achieves up to an 8.56$\times$ speedup over state-of-the-art baselines, delivering low-latency, I/O-efficient long-context inference without sacrificing model accuracy.

He Sun, Jinrui Zhou, Li Li et al.

Large Language Models (LLMs) perform well on many NLP tasks, but fine-tuning them on resource-constrained mobile devices is challenging due to high memory and computation costs, despite growing demands for privacy-preserving personalization. Federated Learning (FL) enables local-data training, yet existing methods either rely on memory-intensive backpropagation or use zeroth-order optimization (ZOO), which avoids backward passes but suffers from slow convergence and degraded accuracy. We propose CooperLLM, a cloud-assisted edge-end cooperative federated fine-tuning framework that combines ZOO on mobile devices with cloud-guided gradient rectification. Mobile clients perform lightweight ZOO updates on private data, while the cloud fine-tunes on auxiliary public data using backpropagation and injects guided perturbations to rectify local updates, improving convergence and accuracy without violating privacy. To address system bottlenecks, CooperLLM introduces pipeline scheduling and adaptive compression to overlap computation and communication and reduce memory usage. Experiments on multiple Transformer models and datasets show that CooperLLM reduces on-device memory by up to $86.4\%$, accelerates convergence by $8.8 \times$, and improves accuracy by up to 10 percentage points over state-of-the-art ZOO-based baselines.