Sparse attention can effectively alleviate the significant demands on memory when large language models (LLMs) process long contexts. Existing methods typically apply the same sparse pattern across different attention heads and inputs. However, this uniform approach fails to capture the inherent diversity of attention patterns within LLMs — the intrinsic attention clustering. To address this, we propose ClusterAttn, a training-free sparse attention method that provides an efficient prompt cache compression scheme under intrinsic attention clustering for efficient LLM inference.Our findings show that attention heads consistently focus on specific clusters of the prompt during decoding, a pattern detectable from an observation window at the prompt’s end. ClusterAttn adaptively fits these clusters utilizing a density-based attention clustering algorithm, thus compressing the KV cache of the prompt. Evaluations on different models across various benchmarks demonstrate ClusterAttn’s superior compression rates and efficiency. By utilizing only 1024 tokens, it can reduce memory usage by 10%–65%, resulting in a latency reduction of 12%–23% and a throughput increase of 2.6–4.8 times, all with nearly no accuracy loss. Additionally, ClusterAttn can handle up to 128k context on a single A100-80GB GPU, outperforming existing methods.

Federated learning (FL) has recently been applied to the parameter-efficient fine-tuning (PEFT) of large language models (LLMs). While promising, client resource heterogeneity has imposed the challenge of the "bucket effect" to FL, where model configuration must cater to the client with the fewest resources. To tackle this issue, heterogeneous low-rank adaptation (LoRA) has recently emerged in FL, which enables clients to do local fine-tuning with different LoRA ranks. However, existing works in this area typically adopt zero-padding, stacking, or singular value decomposition (SVD) for LoRA aggregation, which often incur precision loss or significant overhead, limiting their practicality. In this article, we propose ECLoRA, a novel method for federated fine-tuning with heterogeneous LoRA settings across clients. ECLoRA employs randomized SVD (RSVD) to dramatically reduce aggregation overhead while introducing an error compensation (EC) mechanism that incorporates the decomposition error from previous rounds to improve aggregation precision. Extensive experiments on four widely used foundation models across six public tasks demonstrate the effectiveness of ECLoRA. Specifically, ECLoRA is: (1) accurate, significantly improving the final model performance; (2) fast, accelerating convergence with an average speedup of $1.54\times $ to $3.01\times $ ; and (3) practical, reducing aggregation time by approximately $40\times $ compared to classical SVD.

Pre-trained foundation models, particularly large language models, have achieved remarkable success and led to massive fine-tuned variants. These models are commonly fine-tuned locally and then uploaded by users to cloud platforms such as HuggingFace for secure storage. However, the huge model number and their billion-level parameters impose heavy storage overhead for cloud with limited resources. Our empirical and theoretical analysis reveals that most fine-tuned models in cloud have a small difference (delta) from their pre-trained models. To this end, we propose a novel lossless compression scheme FM-Delta specifically for storing massive fine-tuned models in cloud. FM-Delta maps fine-tuned and pre-trained model parameters into integers with the same bits, and entropy codes their integer delta. In this way, cloud only needs to store one uncompressed pre-trained model and other compressed fine-tuned models. Extensive experiments have demonstrated that FM-Delta efficiently reduces cloud storage consumption for massive fine-tuned models by an average of around 50% with only negligible additional time in most end-to-end cases. For example, on up to 10 fine-tuned models in the GPT-NeoX-20B family, FM-Delta reduces the original storage requirement from 423GB to 205GB, significantly saving cloud storage costs.

Machine learning-based services offer intelligent solutions with powerful models. To enhance model robustness, Federated Learning (FL) emerges as a promising collaborative learning paradigm, which iteratively trains a global model through parameter exchange among multiple clients based on their local data. Generally, the local data are heterogeneous, which slows down convergence. Knowledge distillation is an effective technique against data heterogeneity while existing works distill the ensemble knowledge from local models, ignoring the natural global knowledge from the aggregated model. This places limitations on their algorithms, such as the need for proxy data or the necessary exposure of local models to the server, which is prohibited in most privacy-preserving FL with a clueless server. In this work, we propose FedDGT, a novel knowledge distillation method for industrial server-clueless FL. FedDGT regards the aggregated model as the only one teacher to impart its global knowledge into a generator and then regularizes the drifted local models through the generator, overcoming previous limitations and providing better privacy and scalability support. Extensive experiments demonstrate that FedDGT can achieve highly-competitive model performance while greatly reducing the communication rounds in a server-clueless scenario.

Distributed edge learning collaborates powerful edge devices to train a shared global model. Since the frequent communication between the server and workers is very expensive, it is desired to accelerate the learning process. The gradient sparsification is an efficient method that only uploads a small subset of gradient elements. However, most existing works neglect the distributed nature of local datasets, and consequently the local gradients uploaded by edge devices cannot follow the global correct optimization direction well, which results in the loss of accuracy. In this paper, we propose a new gradient sparsification with a renovating mechanism, called Global Renovating Stochastic Gradient Descent (GRSGD). GRSGD utilizes the previous-round global gradient to estimate the current global one and renovates the current zero-sparsified local gradients. It mitigates the communication overhead while making the convergence direction of training closer to the global optimization, accelerating the distributed edge learning process. We provide a theoretical convergence guarantee for our algorithm based on the non-convex assumption, which better fits most deep learning problems. With extensive experiments in PyTorch, we show that GRSGD effectively accelerates the learning process with a smaller communication cost and a faster convergence rate on most training tasks. For example, on ImageNet MnasNet, GRSGD cuts down the gradient size from 8.47MB to 2.13MB while achieving 9.6%+ higher accuracy.