Scholar Feed

Mixture-of-experts routing

EdgeMoE

Edge-MoE: Memory-Efficient Multi-Task Vision Transformer Architecture with Task-level Sparsity via Mixture-of-Experts

Superseded baseline#16 of 1,370 most-superseded · first seen May 30, 2023

Superseded — cited as a baseline and beaten by newer methods

5 papers critique it · 1 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites EdgeMoE as a baseline.

EdgeMoE's static approach determines optimal bit widths based on specific dataset profiling, leading to inflexibility across diverse environments and potential accuracy impacts.
— HOBBIT: A Mixed Precision Expert Offloading System for Fast MoE Inference
“However, this method does not match the dynamic activation characteristic of MoE models.”
— Fate: Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate
“The majority of existing methods are designed for relatively stable resource environments and generally fail to account for the dynamic heterogeneity of edge network resource states and varying application requirements.”
— CoMoE: Collaborative Optimization of Expert Aggregation and Offloading for MoE-based LLMs at Edge
“While EdgeMoE demonstrates comparable perplexity to D²MoE, it exhibits reduced accuracy in specific zero-shot tasks. This disparity can be ascribed to the differing significance attributed to the various markers, with EdgeMoE utilizing a predetermined mixture of bit-width to ascertain the importance of the experts, resulting in diminished accuracy.”
— D$^{2}$MoE: Dual Routing and Dynamic Scheduling for Efficient On-Device MoE-based LLM Serving
“While static mixed-precision (e.g., EdgeMoE) addresses this, it remains oblivious to dynamic input complexity at runtime”
— DyMoE: Dynamic Expert Orchestration with Mixed-Precision Quantization for Efficient MoE Inference on Edge

Beaten on benchmarks

Head-to-head results where a newer method reports beating EdgeMoE. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.