Zheng Lin

Yuxin Zhang, Zheng Lin, Zhe Chen et al.

Traditional federated learning (FL) frameworks rely heavily on terrestrial networks, where coverage limitations and increasing bandwidth congestion significantly hinder model convergence. Fortunately, the advancement of low-Earth orbit (LEO) satellite networks offers promising new communication avenues to augment traditional terrestrial FL. Despite this potential, the limited satellite-ground communication bandwidth and the heterogeneous operating environments of ground devices-including variations in data, bandwidth, and computing power-pose substantial challenges for effective and robust satellite-assisted FL. To address these challenges, we propose SatFed, a resource-efficient satellite-assisted heterogeneous FL framework. SatFed implements freshness-based model prioritization queues to optimize the use of highly constrained satellite-ground bandwidth, ensuring the transmission of the most critical models. Additionally, a multigraph is constructed to capture real-time heterogeneous relationships between devices, including data distribution, terrestrial bandwidth, and computing capability. This multigraph enables SatFed to aggregate satellite-transmitted models into peer guidance, enhancing local training in heterogeneous environments. Extensive experiments with real-world LEO satellite networks demonstrate that SatFed achieves superior performance and robustness compared to state-of-the-art benchmarks.

Zheng Lin, Yuxin Zhang, Zhe Chen et al.

Recently, the increasing deployment of LEO satellite systems has enabled various space analytics (e.g., crop and climate monitoring), which heavily relies on the advancements in deep learning (DL). However, the intermittent connectivity between LEO satellites and ground station (GS) significantly hinders the timely transmission of raw data to GS for centralized learning, while the scaled-up DL models hamper distributed learning on resource-constrained LEO satellites. Though split learning (SL) can be a potential solution to these problems by partitioning a model and offloading primary training workload to GS, the labor-intensive labeling process remains an obstacle, with intermittent connectivity and data heterogeneity being other challenges. In this paper, we propose LEO-Split, a semi-supervised (SS) SL design tailored for satellite networks to combat these challenges. Leveraging SS learning to handle (labeled) data scarcity, we construct an auxiliary model to tackle the training failure of the satellite-GS non-contact time. Moreover, we propose a pseudo-labeling algorithm to rectify data imbalances across satellites. Lastly, an adaptive activation interpolation scheme is devised to prevent the overfitting of server-side sub-model training at GS. Extensive experiments with real-world LEO satellite traces (e.g., Starlink) demonstrate that our LEO-Split framework achieves superior performance compared to state-ofthe-art benchmarks.

Zehang Lin, Zheng Lin, Miao Yang et al.

The increasing complexity of neural networks poses a significant barrier to the deployment of distributed machine learning (ML) on resource-constrained devices, such as federated learning (FL). Split learning (SL) offers a promising solution by offloading the primary computing load from edge devices to a server via model partitioning. However, as the number of participating devices increases, the transmission of excessive smashed data (i.e., activations and gradients) becomes a major bottleneck for SL, slowing down the model training. To tackle this challenge, we propose a communication-efficient SL framework, named SL-ACC, which comprises two key components: adaptive channel importance identification (ACII) and channel grouping compression (CGC). ACII first identifies the contribution of each channel in the smashed data to model training using Shannon entropy. Following this, CGC groups the channels based on their entropy and performs group-wise adaptive compression to shrink the transmission volume without compromising training accuracy. Extensive experiments across various datasets validate that our proposed SL-ACC framework takes considerably less time to achieve a target accuracy than state-of-the-art benchmarks.

Dingyu Yao, Chenxu Yang, Zhengyang Tong et al.

The Key-Value (KV) cache introduces substantial memory overhead during large language model (LLM) inference. Although existing vector quantization (VQ) methods reduce KV cache usage and provide flexible representational capacity across bit-widths, they suffer severe performance degradation at ultra-low bit-widths due to key cache outliers that hinder effective codebook utilization. To address this challenge, we propose VecInfer, a novel VQ method for aggressive KV cache compression while enabling efficient inference. By applying smooth and Hadamard transformations, VecInfer suppresses outliers in the key cache, enabling the codebook to comprehensively cover the original data distribution and thereby reducing quantization difficulty. To facilitate efficient deployment, we design an optimized CUDA kernel that fuses computation with dequantization to minimize memory access overhead. Extensive evaluations demonstrate that VecInfer consistently outperforms existing quantization baselines across both long-context understanding and mathematical reasoning tasks. With only 2-bit quantization, VecInfer achieves performance comparable to full precision, while delivering up to $\mathbf{2.7\times}$ speedup in large-batch self-attention computation and $\mathbf{8.3\times}$ reduction in single-batch end-to-end latency on Llama-3.1-8B with a 196k sequence length.

Dingyu Yao, Bowen Shen, Zheng Lin et al.

The Key-Value (KV) cache in generative large language models (LLMs) introduces substantial memory overhead. Existing works mitigate this burden by offloading or compressing the KV cache. However, loading the entire cache incurs significant latency due to PCIe bandwidth bottlenecks in CPU-GPU communication, while aggressive compression causes notable performance degradation. We identify that certain layers in the LLM need to maintain global information and are unsuitable for selective loading. In contrast, other layers primarily focus on a few tokens with dominant activations that potentially incur substantial quantization error. This observation leads to a key insight that loading dominant tokens and quantizing all tokens can complement each other. Building on this insight, we propose a hybrid compression method, TailorKV, which seamlessly integrates quantization and offloading. TailorKV develops an inference framework along with a hardware-friendly implementation that leverages these complementary characteristics. Extensive long-context evaluations exhibit that TailorKV achieves nearly lossless performance under aggressive compression settings, outperforming the state-of-the-art. Particularly, the Llama-3.1-8B with 128k context can be served within a single RTX 3090 GPU, reaching 82 ms per token during decoding.