The $qs$ Inequality: Quantifying the Double Penalty of Mixture-of-Experts at Inference

• URL: http://arxiv.org/abs/2603.08960v1
• Date: Mon, 09 Mar 2026 21:48:04 GMT
• Title: The $qs$ Inequality: Quantifying the Double Penalty of Mixture-of-Experts at Inference
• Abstract summary: Mixture-of-Experts (MoE) models deliver high quality at low training FLOPs, but this efficiency often vanishes at inference.<n>We identify a double penalty that structurally disadvantages MoE architectures during decoding.<n>We introduce the $qs$ inequality, a predictive criterion that identifies when MoE is structurally disadvantaged relative to a quality-matched dense model.
• Score: 0.8986240899504297
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Mixture-of-Experts (MoE) models deliver high quality at low training FLOPs, but this efficiency often vanishes at inference. We identify a double penalty that structurally disadvantages MoE architectures during decoding: first, expert routing fragments microbatches and reduces weight reuse; second, massive resident expert pools reduce high-bandwidth memory (HBM) headroom for the KV cache. This phenomenon, formalized as reuse fragmentation, pushes feed-forward networks (FFNs) into a bandwidth-bound regime, especially at long context lengths. We introduce the $qs$ inequality, a predictive criterion that identifies when MoE is structurally disadvantaged relative to a quality-matched dense model. This criterion unifies sparsity ($s$), the fraction of parameters activated per token, and the quality-equivalence factor ($q$), the size multiplier required for a dense model to match MoE performance. Our evaluation across frontier models including DeepSeek-V3, Qwen3-235B, Grok-1, and Switch-C demonstrates that this fragmentation is a general architectural phenomenon. For DeepSeek-V3 at 128k context, this results in a 4.5x throughput advantage for a quality-matched dense baseline. Crucially, massive architectures like Switch-C can become infeasible on cluster sizes where a quality-matched dense model remains viable. Our results suggest that training-time FLOP efficiency is an incomplete proxy for inference-time performance in long-context serving. They also indicate that MoE may be best viewed as a training-time optimization, with distillation into dense models as a possible path toward inference-efficient deployment.

Related papers

• Variable-Width Transformers [57.07167443557886]
  We investigate nonuniform capacity allocation across network depth by proposing a $times$-shaped > former architecture.<n>Our design maintains wider early and late layers while narrowing the middle layers, utilizing a parameter-free residual resizing mechanism.<n>By reducing the average layer width, this architecture also requires fewer overall FLOPs and smaller KV cache memory and I/O cost.
  arXiv  Detail & Related papers  (2026-06-16T17:59:03Z)
• OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond [50.440302567029654]
  Multi-modal intelligence has made the Key-Value cache a dominant memory bottleneck for efficient deployment.<n>In this work, we revisit the inherent limitations of the per-channel quantization paradigm.<n>We propose OScaR, an accurate and lightweight KV cache compression framework for X-LLMs.
  arXiv  Detail & Related papers  (2026-05-19T10:53:03Z)
• Machine Learning Models to Identify Promising Nested Antiresonance Nodeless Fiber Designs [0.0]
  We propose a machine learning framework designed to identify high-performance nested antiresonance nodeless fibers (NANFs)<n>We show that small data sets are sufficient for stable, high-accuracy performance prediction, enabling the exploration of design spaces as large as $14e6$ cases.
  arXiv  Detail & Related papers  (2026-03-02T14:13:59Z)
• Mixture-of-Experts Models in Vision: Routing, Optimization, and Generalization [0.0]
  We study MoE behavior in an image classification setting, focusing on predictive performance, expert utilization, and generalization.<n>We compare dense, SoftMoE, and SparseMoE classifier heads on the CIFAR10 dataset under comparable model capacity.<n>Both MoE variants achieve slightly higher validation accuracy than the dense baseline while maintaining balanced expert utilization through regularization.<n>We find that SoftMoE exhibits higher sharpness by these metrics, while Dense and SparseMoE lie in a similar curvature regime, despite all models achieving comparable generalization performance.
  arXiv  Detail & Related papers  (2026-01-21T14:22:25Z)
• Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2 [0.0]
  We show that MAW-guided width pruning acts as a selective filter, reducing parametric knowledge while preserving or enhancing behavioral alignment.<n>We quantify context-dependent efficiency trade-offs: pruned configurations achieve up to 23% reduction in energy consumption (J/token) but incur penalties in single-request latency.
  arXiv  Detail & Related papers  (2025-12-27T18:09:57Z)
• Breaking the MoE LLM Trilemma: Dynamic Expert Clustering with Structured Compression [14.086434595924716]
  Mixture-of-Experts (MoE) Large Language Models (LLMs) face a trilemma of load imbalance, parameter redundancy, and communication overhead.<n>We introduce a unified framework based on dynamic expert clustering and structured compression to address these issues cohesively.
  arXiv  Detail & Related papers  (2025-09-27T10:45:58Z)
• PT$^2$-LLM: Post-Training Ternarization for Large Language Models [52.4629647715623]
  Large Language Models (LLMs) have shown impressive capabilities across diverse tasks, but their large memory and compute demands hinder deployment.<n>We propose PT$2$-LLM, a post-training ternarization framework tailored for LLMs.<n>At its core is an Asymmetric Ternary Quantizer equipped with a two-stage refinement pipeline.
  arXiv  Detail & Related papers  (2025-09-27T03:01:48Z)
• Scaling Linear Attention with Sparse State Expansion [62.749291436866606]
  Transformer architecture struggles with long-context scenarios due to quadratic computation and linear memory growth.<n>We propose two key innovations to achieve more effective context compression.<n>First, we introduce a row-sparse update formulation for linear attention by conceptualizing state updating as information classification.<n>Second, we present Sparse State Expansion (SSE) within the sparse framework, which expands the contextual state into multiple partitions.
  arXiv  Detail & Related papers  (2025-07-22T13:27:31Z)
• MPQ-DMv2: Flexible Residual Mixed Precision Quantization for Low-Bit Diffusion Models with Temporal Distillation [74.34220141721231]
  We present MPQ-DMv2, an improved textbfMixed textbfPrecision textbfQuantization framework for extremely low-bit textbfDiffusion textbfModels.
  arXiv  Detail & Related papers  (2025-07-06T08:16:50Z)
• Flash-LLM: Enabling Cost-Effective and Highly-Efficient Large Generative Model Inference with Unstructured Sparsity [12.663030430488922]
  We propose Flash-LLM for enabling low-cost and highly-efficient large generative model inference on high-performance Cores. At SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9x and 1.5x, respectively.
  arXiv  Detail & Related papers  (2023-09-19T03:20:02Z)
• SqueezeLLM: Dense-and-Sparse Quantization [80.32162537942138]
  Main bottleneck for generative inference with LLMs is memory bandwidth, rather than compute, for single batch inference. We introduce SqueezeLLM, a post-training quantization framework that enables lossless compression to ultra-low precisions of up to 3-bit. Our framework incorporates two novel ideas: (i) sensitivity-based non-uniform quantization, which searches for the optimal bit precision assignment based on second-order information; and (ii) the Dense-and-Sparse decomposition that stores outliers and sensitive weight values in an efficient sparse format.
  arXiv  Detail & Related papers  (2023-06-13T08:57:54Z)
• Rethinking Lightweight Salient Object Detection via Network Depth-Width Tradeoff [26.566339984225756]
  Existing salient object detection methods often adopt deeper and wider networks for better performance. We propose a novel trilateral decoder framework by decoupling the U-shape structure into three complementary branches. We show that our method achieves better efficiency-accuracy balance across five benchmarks.
  arXiv  Detail & Related papers  (2023-01-17T03:43:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.