The paper presents a method for distributing large language model inference across multiple Intel AI PCs by splitting the model into pipeline shards, each pre‑compiled into an OpenVINO graph. Three key techniques—beam_idx Gather to enable GPU optimizations, speculative decoding on stateful models, and interleaved micro‑batching—allow a two‑node Llama 3.1 8B INT4 pipeline to serve two users at 1.79× the throughput of a single‑node model, while a four‑node deployment can run a 70B model that no single PC can hold. The authors provide code, benchmark logs, and reproduction scripts on GitHub.
By Tate Berenbaum, Muthaiah Venkatachalam
arXiv:2606. 00946v1 Announce Type: cross Abstract: Efficiently serving large language model (LLM) inference tasks is crucial both for user-perceived latency such as time-to-first-token (TTFT) and for GPU utilization.
By Gangmuk Lim, Wanyu Zhao, Brighten Godfrey, Jiaxin Shan, Le Xu, Liguang Xie
arXiv:2602. 24044v2 Announce Type: replace-cross Abstract: Large Language Model (LLM) adapters enable low-cost model specialization, but introduce complex caching and scheduling challenges in distributed serving systems where hundreds of adapters must be hosted concurrently.
By Ferran Agullo, Joan Oliveras, Chen Wang, Alberto Gutierrez-Torre, Olivier Tardieu, Alaa Youssef, Jordi Torres, Josep Ll. Berral
The paper proposes two hybrid algorithms, FL+FSDP and FL+HSDP, that combine sharded data parallelism with federated learning-style aggregations to reduce communication overhead in large-scale AI training. By forming loosely‑coupled federation groups, the methods keep inter‑group traffic minimal while maintaining a bounded global batch size. Experiments on a Llama3.1 8B model trained on 512 A100 GPUs show up to 8.04× faster data processing and 4.48 lower evaluation perplexity compared to traditional sharded DP approaches.
The paper proposes two hybrid algorithms, FL+FSDP and FL+HSDP, that combine sharded data parallelism with federated learning-style aggregations to reduce communication overhead in large-scale AI training. By partitioning GPUs into loosely‑coupled federation groups, the methods keep inter‑group traffic minimal while maintaining a bounded global batch size. Experiments on a Llama3.1 8B model trained on 512 A100 GPUs show up to 8.04× faster data processing and 4.48 lower evaluation perplexity compared to conventional sharded DP.
By Gianluca Mittone, Marco Aldinucci
The paper introduces the Power Flexibility Index (PFI) to measure how large language model (LLM) training performance changes when GPU power is reduced. Using 131 training runs on H200 and H100 GPUs, the study finds that LLM jobs have significant but variable power elasticity and identifies telemetry signals that can predict PFI during runtime. The authors demonstrate that allocating power based on PFI maximizes overall token throughput, recovering about 1.5k tokens/s per job under a 30% power reduction, which represents 63% of the gap between equal-weight and perfect-information allocations.
By Philip Colangelo, Charles Dawson, Shayan Sengupta, Ayse Coskun, Varun Sivaram