Serving a 235B-parameter Mixture-of-Experts (MoE) model on a single 8 GB GPU is bottlenecked not by compute but by memory bandwidth: decode must stream each token's active experts from whichever tier holds them, and on consumer hardware most experts sit on an SSD far slower than RAM. We quantify this bandwidth wall on Qwen3-235B (Q4_K_M, 134 GB): measured decode is 0.
The paper introduces a cache‑aware post‑training framework for Mixture‑of‑Experts (MoE) models that jointly adapts the MoE backbone and lightweight auxiliary cache routers while keeping the native Top‑K expert‑selection rule. Two modes are proposed: Temporal Router, which predicts same‑layer reuse and retains experts for future tokens, and Spatio‑Temporal Router, which adds a Spatio Router that refines the temporal cache using the causal predecessor’s hidden state. Experiments on Qwen3 and GPT‑OSS across GSM8K, MATH, and CommonsenseQA show that Temporal Router improves cache hit rates and reduces expert‑weight traffic, while Spatio‑Temporal Router achieves the best load‑adjusted efficiency, outperforming strong prefetching baselines.
By Zhenhe Wu, Yaping Jin, Qinghua Xing, Hang Zhou, Wei He, Xianjie Wu, Xianfu Cheng, Jian Yang, Hanting Chen
WiSP (Working‑Set Paging) is a routing‑aware expert pager that allows Mixture‑of‑Experts models to run on GPUs that cannot hold the entire expert pool by paging experts in and out of VRAM while preserving byte‑identical outputs. On a 24 GiB RTX 3090, WiSP doubles decode throughput compared to static offload when the model does not fit, and its companion policy MV‑WSA allocates VRAM between resident experts and KV cache based on marginal latency benefit, reducing end‑to‑end time by up to 1.19× without altering model outputs.
By Jiamu Zhang, Liang Wu, Mayank Darbari, Liangjie Hong
Dynamic Expert Quantization (DynaExq) is a runtime-aware mixed-precision serving system designed for single‑GPU Mixture‑of‑Experts (MoE) inference under a hard high‑bandwidth memory (HBM) envelope. It treats the problem as an online, budget‑constrained precision allocation task, keeping the most frequently used experts at higher precision while relegating the rest to low‑precision fallbacks. By estimating expert hotness from router traces and asynchronously promoting or demoting experts, DynaExq maintains a fully materialized expert set during the forward pass, improving accuracy and throughput compared to static post‑training quantization and offloading/prefetch baselines.
whyItMatters":"DynaExq enables efficient deployment of large MoE models on memory‑limited GPUs by dynamically allocating precision based on runtime expert usage, thereby reducing memory footprint and latency while boosting accuracy and throughput."
By Kexin Chu, Dawei Xiang, Zixu Shen, Yiwei Yang, Zecheng Liu, Wei Zhang
arXiv:2604. 26968v2 Announce Type: replace-cross Abstract: Key-value (KV) cache memory management is the primary bottleneck limiting throughput and cost-efficiency in large-scale GPU inference serving.
By Sanjeev Rao Ganjihal
The paper argues that for large‑context autoregressive language‑model inference, memory bandwidth—specifically the Key‑Value (KV) cache—becomes the limiting resource rather than arithmetic throughput. It analytically derives how arithmetic intensity decays with context length for NVIDIA H100, NVIDIA B200, and AMD MI300X, identifies crossover points where KV traffic overtakes weight traffic, and evaluates representative techniques across five compression domains. The study finds a three‑regime behavior: below the crossover, weight traffic dominates and KV compression offers little benefit; beyond it, KV traffic dominates and compression methods trade quality for bandwidth, with paging and prefix sharing being lossless but capacity‑limited, while quantization and eviction directly reduce bandwidth at the cost of accuracy.
whyItMatters":"The work provides a unified analytical framework and a standardized protocol that enable consistent comparison of KV‑compression techniques across hardware and workloads, guiding practitioners in selecting appropriate methods for long‑context inference."
By Tejinder Singh