Efficient inference of large language models (LLMs) is hindered by an ever-growing key-value (KV) cache, making KV cache compression a critical research direction. Traditional methods selectively evict less important KV cache entries, which leads to information loss and hallucinations. Recently, merging-based strategies have been explored to retain more information by merging KV pairs that would be discarded; however, these existing approaches inevitably introduce inconsistencies in attention distributions before and after merging, causing degraded generation quality. To overcome this challenge, we propose KeepKV , a novel adaptive KV cache merging method designed to preserve performance under strict memory constraints, achieving single-step lossless compression and providing error bounds for multi-step compression. KeepKV introduces the Electoral Votes mechanism that records merging history and adaptively adjusts attention scores. Moreover, it further leverages a novel Zero Inference-Perturbation Merging method, compensating for attention loss resulting from cache merging. Extensive experiments on various benchmarks and LLM architectures demonstrate that KeepKV substantially reduces memory usage while successfully retaining essential context information, achieving over 2 times inference throughput improvement and maintaining superior generation quality even with only 10% KV cache budgets.

Model merging has emerged as a promising technique for enhancing large language models, though its application in large-scale pre-training remains relatively unexplored. In this paper, we present a comprehensive investigation of model merging techniques during the pre-training process. Through extensive experiments with both dense and Mixture-of-Experts (MoE) architectures ranging from millions to over 100 billion parameters, we demonstrate that merging checkpoints trained with constant learning rates not only achieves significant performance improvements but also enables accurate prediction of annealing behavior. These improvements lead to both more efficient model development and significantly lower training costs. Our detailed ablation studies on merging strategies and hyperparameters provide new insights into the underlying mechanisms while uncovering novel applications. Through comprehensive experimental analysis, we offer the open-source community practical pre-training guidelines for effective model merging.

Long tail latency of short flows (or messages) greatly affects user-facing applications in datacenters. Prior solutions to the problem introduce significant implementation complexities, such as global state monitoring, complex network control, or non-trivial switch modifications. While promising superior performance, they are hard to implement in practice.This paper presents Cloudburst, a simple, effective yet readily deployable solution achieving similar or even better results without introducing the above complexities. At its core, Cloudburst explores forward error correction (FEC) over multipath — it proactively spreads FEC-coded packets generated from messages over multipath in parallel, and recovers them with the first few arriving ones. As a result, Cloudburst is able to obliviously exploit underutilized paths, thus achieving low tail latency. We have implemented Cloudburst as a user-space library, and deployed it on a testbed with commodity switches. Our testbed and simulation experiments show the superior performance of Cloudburst. For example, Cloudburst achieves 63.69% and 60.06% reduction in 99th percentile message/flow completion time (FCT) compared to DCTCP and PIAS, respectively.

We present ByteScheduler, a generic communication scheduler for distributed DNN training acceleration. ByteScheduler is based on our principled analysis that partitioning and rearranging the tensor transmissions can result in optimal results in theory and good performance in real-world even with scheduling overhead. To make ByteScheduler work generally for various DNN training frameworks, we introduce a unified abstraction and a Dependency Proxy mechanism to enable communication scheduling without breaking the original dependencies in framework engines. We further introduce a Bayesian Optimization approach to auto-tune tensor partition size and other parameters for different training models under various networking conditions. ByteScheduler now supports TensorFlow, PyTorch, and MXNet without modifying their source code, and works well with both Parameter Server (PS) and all-reduce architectures for gradient synchronization, using either TCP or RDMA. Our experiments show that ByteScheduler accelerates training with all experimented system configurations and DNN models, by up to 196% (or 2.96X of original speed).

Leveraging application-level requirements using coflows has recently been shown to improve application-level communication performance in data-parallel clusters. However, existing coflow-based solutions rely on modifying applications to extract coflows, making them inapplicable to many practical scenarios.