arXiv:2608. 12385v2 Announce Type: replace Abstract: As large language models serve ever more requests, cumulative inference cost is growing relative to the one-time cost of training.
The paper introduces Cascadia, a resident inference system that runs the 975B‑parameter Inkling mixture‑of‑experts model across eleven Intel Core Ultra X7 358H AI PCs. It details a custom MoE engine that preserves routing rules, compresses graphs for OpenVINO, and coordinates FP16 expert computation with FP32 output restoration, enabling efficient deployment of six decoder layers per machine. Performance results show up to 60.29 aggregate decode tokens per second at 88 streams, with first‑token latency of 6.05 s at fifteen streams, and demonstrate that larger context budgets recover embedded code while maintaining bounded latency.
arXiv:2609.14507v1 Announce Type: cross Abstract: Single-GPU long-context inference with Mixture-of-Experts (MoE) models requires spilling the key-value cache (KVCache) to CPU memory. The spilled KV...
arXiv:2607. 23264v1 Announce Type: cross Abstract: Fine-grained, device-initiated communication lets persistent GPU kernels in distributed diffusion transformer (DiT) inference issue remote stores and overlap data movement with Tensor Core computation.
arXiv:2606. 10493v1 Announce Type: cross Abstract: Local deployment of large Mixture-of-Experts (MoE) models falls short of the service quality achieved in cloud-scale environments, even under low-concurrency workloads.
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.
The paper investigates the performance impact of using a remote KV cache tier for long‑context language model serving. In a disaggregated deployment of GLM‑5.3 on 8xB300 nodes, enabling the remote tier initially increased the 16k‑token p99 latency from 3.96 s to 17.3 s, and synchronous loading caused severe queue instability. By sharding the KV store across context‑parallel ranks, partitioning the remote tier, adding asynchronous loading, and fixing client bugs, the authors reduced the p99 latency from 83.32 s to 0.67 s and increased prefill throughput by 14–34 %. whyItMatters":"The study demonstrates that careful sharding and asynchronous handling of remote KV caches can dramatically improve latency and throughput in large‑scale LLM serving systems."
FluxMoE introduces an expert paging system that decouples Mixture-of-Experts (MoE) model experts from permanent GPU residency, allowing dynamic adaptation to available memory. By combining PagedTensor, a bandwidth‑balanced memory hierarchy, and a budget‑aware residency planner, FluxMoE streams expert weights on demand while keeping computations on GPUs. Experiments on GLM‑4.5 and Mixtral‑8×7B‑Instruct show significant throughput gains and reduced time‑per‑output‑token compared to existing inference engines, without compromising model quality.
arXiv:2607. 05876v1 Announce Type: cross Abstract: LLM serving optimization typically benchmarks many configurations and reaches for heavy profilers when latency targets are missed.
arXiv:2607. 22785v1 Announce Type: cross Abstract: Apple-Silicon SoCs share CPU, GPU, and Neural Engine over one unified memory system, raising the question of whether transformer inference can be accelerated by splitting single operators across units.
The paper investigates whether training Mixture-of-Experts (MoE) routers can improve memory‑bandwidth locality on consumer GPUs. Using a new zero‑surgery telemetry tool, the authors measure that a large Qwen3‑235B model is bottlenecked by disk‑based expert access, and that an LRU cache can serve a majority of requests. They pre‑register experiments training 137 M‑parameter MoE models with locality‑aware losses, finding that while cache misses can drop up to 60 % (99 % static‑pin hit rate), every configuration fails to meet a strict 1 % perplexity threshold, indicating a tight coupling between cache efficiency and model quality.
DualPace is an LLM serving runtime that decouples scheduling between a GPU engine and an AMX‑enabled CPU engine, allowing the CPU to independently batch and advance decodes while still cooperating with the GPU for dense stages. This design expands key‑value cache capacity into host memory and reduces bursty CPU progress, leading to significant throughput gains—up to 43.2 % over NEO and 90.6 % over vLLM—without increasing latency. The system also maintains low time‑to‑first‑token and supports higher request rates at comparable tail latency.