On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention

Large language models (LLMs) excel at capturing global token dependencies via self-attention but face prohibitive compute and memory costs on lengthy inputs. While sub-quadratic methods (e.g., linear attention) can reduce these costs, they often degrade accuracy due to overemphasizing recent tokens. In this work, we first propose dual-state linear attention (DSLA), a novel design that maintains two specialized hidden states-one for preserving historical context and one for tracking recency-thereby mitigating the short-range bias typical of linear-attention architectures. To further balance efficiency and accuracy under dynamic workload conditions, we introduce DSLA-Serve, an online adaptive distillation framework that progressively replaces Transformer layers with DSLA layers at inference time, guided by a sensitivity-based layer ordering. DSLA-Serve uses a chained fine-tuning strategy to ensure that each newly converted DSLA layer remains consistent with previously replaced layers, preserving the overall quality. Extensive evaluations on commonsense reasoning, long-context QA, and text summarization demonstrate that DSLA-Serve yields 2.3x faster inference than Llama2-7B and 3.0x faster than the hybrid Zamba-7B, while retaining comparable performance across downstream tasks. Our ablation studies show that DSLA's dual states capture both global and local dependencies, addressing the historical-token underrepresentation seen in prior linear attentions. Codes are available at https://github.com/utnslab/DSLA-Serve.

HALoS: Hierarchical Asynchronous Local SGD over Slow Networks for Geo-Distributed Large Language Model Training

Training large language models (LLMs) increasingly relies on geographically distributed accelerators, causing prohibitive communication costs across regions and uneven utilization of heterogeneous hardware. We propose HALoS, a hierarchical asynchronous optimization framework that tackles these issues by introducing local parameter servers (LPSs) within each region and a global parameter server (GPS) that merges updates across regions. This hierarchical design minimizes expensive inter-region communication, reduces straggler effects, and leverages fast intra-region links. We provide a rigorous convergence analysis for HALoS under non-convex objectives, including theoretical guarantees on the role of hierarchical momentum in asynchronous training. Empirically, HALoS attains up to 7.5x faster convergence than synchronous baselines in geo-distributed LLM training and improves upon existing asynchronous methods by up to 2.1x. Crucially, HALoS preserves the model quality of fully synchronous SGD-matching or exceeding accuracy on standard language modeling and downstream benchmarks-while substantially lowering total training time. These results demonstrate that hierarchical, server-side update accumulation and global model merging are powerful tools for scalable, efficient training of new-era LLMs in heterogeneous, geo-distributed environments.

ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments

The growing use of autonomous mobile service robots (AMSRs) in dynamic environments requires flexible management of compute resources to optimize the performance of diverse tasks such as navigation, localization, perception, and so on. Current robot deployments, which oftentimes rely on static configurations (of the OS, applications, etc.) and system over-provisioning, fall short since they do not account for the tasks' performance variations resulting in poor system-wide behavior such as robot instability and/or inefficient resource use. This paper presents ConfigBot, a system designed to adaptively reconfigure AMSR applications to meet a predefined performance specification by leveraging runtime profiling and automated configuration tuning. Through experiments on a Boston Dynamics Spot robot equipped with NVIDIA AGX Orin, we demonstrate ConfigBot's efficacy in maintaining system stability and optimizing resource allocation across diverse scenarios. Our findings highlight the promise of tailored and dynamic configurations for robot deployments.

OMEGA: A Low-Latency GNN Serving System for Large Graphs

Graph Neural Networks (GNNs) have been widely adopted for their ability to compute expressive node representations in graph datasets. However, serving GNNs on large graphs is challenging due to the high communication, computation, and memory overheads of constructing and executing computation graphs, which represent information flow across large neighborhoods. Existing approximation techniques in training can mitigate the overheads but, in serving, still lead to high latency and/or accuracy loss. To this end, we propose OMEGA, a system that enables low-latency GNN serving for large graphs with minimal accuracy loss through two key ideas. First, OMEGA employs selective recomputation of precomputed embeddings, which allows for reusing precomputed computation subgraphs while selectively recomputing a small fraction to minimize accuracy loss. Second, we develop computation graph parallelism, which reduces communication overhead by parallelizing the creation and execution of computation graphs across machines. Our evaluation with large graph datasets and GNN models shows that OMEGA significantly outperforms state-of-the-art techniques.

TrainMover: Efficient ML Training Live Migration with No Memory Overhead

Machine learning training has emerged as one of the most prominent workloads in modern data centers. These training jobs are large-scale, long-lasting, and tightly coupled, and are often disrupted by various events in the cluster such as failures, maintenance, and job scheduling. To handle these events, we rely on cold migration, where we first checkpoint the entire cluster, replace the related machines, and then restart the training. This approach leads to disruptions to the training jobs, resulting in significant downtime. In this paper, we present TrainMover, a live migration system that enables machine replacement during machine learning training. TrainMover minimizes downtime by leveraging member replacement of collective communication groups and sandbox lazy initialization. Our evaluation demonstrates that TrainMover achieves 16x less downtime compared to all baselines, effectively handling data center events like straggler rebalancing, maintenance, and unexpected failures.

C3: Learning Congestion Controllers with Formal Certificates

Learning-based congestion controllers offer better adaptability compared to traditional heuristic algorithms. However, the inherent unreliability of learning techniques can cause learning-based controllers to behave poorly, creating a need for formal guarantees. While methods for formally verifying learned congestion controllers exist, these methods offer binary feedback that cannot optimize the controller toward better behavior. We improve this state-of-the-art via C3, a new learning framework for congestion control that integrates the concept of formal certification in the learning loop. C3 uses an abstract interpreter that can produce robustness and performance certificates to guide the training process, rewarding models that are robust and performant even on worst-case inputs. Our evaluation demonstrates that unlike state-of-the-art learned controllers, C3-trained controllers provide both adaptability and worst-case reliability across a range of network conditions.

Read-ME: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design

The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and performance. Despite their benefits, MoE models face significant challenges during inference, including inefficient memory management and suboptimal batching, due to misaligned design choices between the model architecture and the system policies. Furthermore, the conventional approach of training MoEs from scratch is increasingly prohibitive in terms of cost. In this paper, we propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models (in contrast to "upcycling" generalist MoEs), avoiding the high costs of ground-up training. Our approach employs activation sparsity to extract experts. To compose experts, we examine the widely-adopted layer-wise router design and show its redundancy, and thus we introduce the pre-gating router decoupled from the MoE backbone that facilitates system-friendly pre-computing and lookahead scheduling, enhancing expert-aware batching and caching. Our codesign therefore addresses critical gaps on both the algorithmic and system fronts, establishing a scalable and efficient alternative for LLM inference in resource-constrained settings. Read-ME outperforms other popular open-source dense models of similar scales, achieving improvements of up to 10.1% on MMLU, and improving mean end-to-end latency up to 6.1%. Codes are available at: https://github.com/VITA-Group/READ-ME.

CONGO: Compressive Online Gradient Optimization with Application to Microservices Management

We address the challenge of online convex optimization where the objective function's gradient exhibits sparsity, indicating that only a small number of dimensions possess non-zero gradients. Our aim is to leverage this sparsity to obtain useful estimates of the objective function's gradient even when the only information available is a limited number of function samples. Our motivation stems from distributed queueing systems like microservices-based applications, characterized by request-response workloads. Here, each request type proceeds through a sequence of microservices to produce a response, and the resource allocation across the collection of microservices is controlled to balance end-to-end latency with resource costs. While the number of microservices is substantial, the latency function primarily reacts to resource changes in a few, rendering the gradient sparse. Our proposed method, CONGO (Compressive Online Gradient Optimization), combines simultaneous perturbation with compressive sensing to estimate gradients. We establish analytical bounds on the requisite number of compressive sensing samples per iteration to maintain bounded bias of gradient estimates, ensuring sub-linear regret. By exploiting sparsity, we reduce the samples required per iteration to match the gradient's sparsity, rather than the problem's original dimensionality. Numerical experiments and real-world microservices benchmarks demonstrate CONGO's superiority over multiple stochastic gradient descent approaches, as it quickly converges to performance comparable to policies pre-trained with workload awareness.

HawkVision: Low-Latency Modeless Edge AI Serving

The trend of modeless ML inference is increasingly growing in popularity as it hides the complexity of model inference from users and caters to diverse user and application accuracy requirements. Previous work mostly focuses on modeless inference in data centers. To provide low-latency inference, in this paper, we promote modeless inference at the edge. The edge environment introduces additional challenges related to low power consumption, limited device memory, and volatile network environments. To address these challenges, we propose HawkVision, which provides low-latency modeless serving of vision DNNs. HawkVision leverages a two-layer edge-DC architecture that employs confidence scaling to reduce the number of model options while meeting diverse accuracy requirements. It also supports lossy inference under volatile network environments. Our experimental results show that HawkVision outperforms current serving systems by up to 1.6X in P99 latency for providing modeless service. Our FPGA prototype demonstrates similar performance at certain accuracy levels with up to a 3.34X reduction in power consumption.

FFN-SkipLLM: A Hidden Gem for Autoregressive Decoding with Adaptive Feed Forward Skipping

Autoregressive Large Language Models (e.g., LLaMa, GPTs) are omnipresent achieving remarkable success in language understanding and generation. However, such impressive capability typically comes with a substantial model size, which presents significant challenges for autoregressive token-by-token generation. To mitigate computation overload incurred during generation, several early-exit and layer-dropping strategies have been proposed. Despite some promising success due to the redundancy across LLMs layers on metrics like Rough-L/BLUE, our careful knowledge-intensive evaluation unveils issues such as generation collapse, hallucination of wrong facts, and noticeable performance drop even at the trivial exit ratio of 10-15% of layers. We attribute these errors primarily to ineffective handling of the KV cache through state copying during early-exit. In this work, we observed the saturation of computationally expensive feed-forward blocks of LLM layers and proposed FFN-SkipLLM, which is a novel fine-grained skip strategy of autoregressive LLMs. More specifically, FFN-SkipLLM is an input-adaptive feed-forward skipping strategy that can skip 25-30% of FFN blocks of LLMs with marginal change in performance on knowledge-intensive generation tasks without any requirement to handle KV cache. Our extensive experiments and ablation across benchmarks like MT-Bench, Factoid-QA, and variable-length text summarization illustrate how our simple and ease-at-use method can facilitate faster autoregressive decoding.