SYMPHONY: Improving Memory Management for LLM Inference Workloads

Saurabh Agarwal; Anyong Mao; Aditya Akella; Shivaram Venkataraman

Home/Archive/arXiv

arXiv · 2412.16434

SYMPHONY: Improving Memory Management for LLM Inference Workloads

Saurabh Agarwal·Anyong Mao·Aditya Akella·Shivaram Venkataraman

View original source ↗Read source document ↗Export metadata ↓

Abstract

Large Language Models (LLMs) are increasingly being deployed in applications such as chatbots, code editors, and conversational agents. A key feature of LLMs is their ability to engage in multi-turn interactions with humans or external tools, enabling a wide range of tasks. Each new request in a multi-turn interaction depends on the intermediate state, specifically the key-value (K,V) caches, from previous requests in the ongoing interaction. Existing serving engines either recompute the K,V caches or offload them to main memory. Profiling reveals that recomputation can result in over 99% of processed tokens being redundant. On the other hand, offloading K,V caches from GPU memory makes inference serving stateful, leading to load imbalances across the cluster. To address these challenges, we developed SYMPHONY. SYMPHONY leverages the observation that multi-turn work loads provide additional hints that allow K,V caches to be migrated off the critical serving path. By utilizing these hints, SYMPHONY dynamically migrates K,V caches to enable finegrained scheduling of inference requests. Our experiments demonstrate that SYMPHONY can handle over 8x the number of requests compared to state-of-the-art baselines, with a similar latency profile.

Explore related subjects

cs.DC

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Saurabh Agarwal, Anyong Mao, Aditya Akella, Shivaram Venkataraman. 2024-12-21. SYMPHONY: Improving Memory Management for LLM Inference Workloads. https://arxiv.org/abs/2412.16434

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Fast and Robust Information Spreading in the Noisy PULL Model

Efficient information spreading in stochastic multi agent systems is a core challenge when communication is noisy, bandwidth limited, and agents lack global coordination. Yet biological systems, including ant colonies and fish schools, routinely overcome these constraints. A small number of informed individuals can reliably guide large, uncoordinated populations using minimal, noisy signals. Motivated by these observations, we study how reliable information dissemination can be achieved in such bio inspired settings. A population of $n$ agents, each with a binary preference, includes a designated subset of source agents, and the goal is to converge to the majority preference among the sources. In the noisy $\mathcal{PULL}(h)$ model, each agent observes noisy messages from $h$ randomly sampled peers in every round. Prior work shows that convergence requires $Ω(n/h)$ rounds even under favorable conditions. We ask how far we can push simplicity, with no synchronization at the start time and minimal message size, without compromising convergence speed. We present a quasi self-stabilizing protocol using only 2-bit messages that converges from arbitrary initial states despite severe noise and initial asynchrony. It achieves optimal convergence time $O((n/h)\log n)$ with high probability, and in particular $O(\log n)$ time in the snapshot regime $h=Θ(n)$. A key subroutine is an even simpler 1-bit protocol assuming simultaneous start, based on a natural two phase listen then amplify mechanism. Together, our results show that simple, biologically inspired protocols can achieve optimal and robust information dissemination even in highly unreliable and uncoordinated systems.

cs.DC↗

arXiv2026

LOIP:Collaborative Lossless LLM Inference Serving with Offloading-based Pipeline Parallelism on Edge Devices

Providing lossless inference services of LLMs on edge devices remains challenging, especially given the extremely tight memory budgets. The existing offloading techniques inevitably introduce numerous loading bubbles, which further inflate the end-to-end latency of the entire inference pipeline. Meanwhile, dynamically fluctuating network bandwidth and diverse user request patterns pose additional obstacles to efficient lossless inference on edge devices. To address this, we propose LOIP, a collaborative lossless LLM inference system that employs an offloading-based interleaved pipeline parallelism to better overlap model offloading with computing and communicating. Specifically, LOIP first constructs an offloading-aware cost model to characterize inference latency and memory overhead under heterogeneous device capabilities and limited bandwidth. Based on this cost model, LOIP develops a fine-grained allocation scheduler that determines latency-efficient layer partitions across devices while explicitly accounting for offloading overhead, along with a unified memory architecture (UMA)-aware loading optimization using customized CUDA operators to reduce runtime loading overhead. LOIP further designs an online memory adaptation strategy to handle the increasing KV cache pressure and dynamic bandwidth fluctuations during inference. We implement LOIP with 2500+ lines of Python and 500+ lines of C++/CUDA code, and deploy it on five heterogeneous NVIDIA Jetson edge devices for lossless collaborative inference of LLaMA3.3-70B-Instruct. Extensive experiments demonstrate that LOIP achieves 8.8$\times$$\sim$20.3$\times$ speedups over the SOTA baselines under different bandwidth conditions and request patterns without compromising model accuracy.

cs.DC↗

arXiv2026

Fluid Notarization: Verifiable Evolution of Concurrently Edited Structured Documents

Traditional blockchain-based document notarization follows a snapshot-oriented model in which each document revision is represented as an independent state anchored on-chain through a cryptographic reference. While effective for immutable artifacts, this approach becomes inadequate when documents evolve through collaborative editing. Concurrent modifications create divergent document versions that must be reconciled outside the notarization layer, while even minor changes require generating and distributing new document snapshots. Conversely, collaborative replication frameworks such as CRDTs provide deterministic reconciliation of concurrent updates, but do not inherently provide independently verifiable evidence of when contributions were published. This paper introduces Fluid Notarization, a notarization paradigm in which document evolution itself becomes the object of notarization. Rather than certifying isolated states, Fluid Notarization certifies a graph of causally related evolution artifacts generated by a JSON-native delta-CRDT. The proposed model builds upon Melda, which represents document changes as compact, content-addressed deltas linked through causal dependencies. Blockchain notarization is reduced to recording identifiers of these evolution artifacts, while synchronization, reconstruction, and conflict resolution remain entirely off-chain. The resulting architecture combines two complementary guarantees: deterministic convergence provided by the CRDT and independently auditable proof-of-existence, provenance, and publication evidence provided by the blockchain. A prototype implementation and validation scenario based on collaboratively edited electronic health records demonstrate the feasibility of the approach and highlight the advantages of notarizing document evolution rather than successive document snapshots.

cs.DC↗