Research map / Architectures and efficiency / Mixture of experts
ZipMoE: Efficient On-Device MoE Serving via Lossless Compression and Cache-Affinity Scheduling
Yuchen Yang, Yaru Zhao, Pu Yang, Shaowei Wang, Zhi-Hua Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ZipMoE hierarchical cache management instead of LRU/FIFO/Marking eviction (ablation) | ZipMoE · DeepSeekV2-Lite, 10GB, BS=16 | throughput (tokens/s): 4.43 -> 5.18 | Improved |
| add cache pool planning on top of heterogeneous cache pools (ablation) | ZipMoE · DeepSeekV2-Lite, 10GB, BS=16 | throughput (tokens/s): 5.18 -> 5.30 | Improved |
| add cache pool planning on top of heterogeneous cache pools (ablation) | ZipMoE · Qwen1.5-MoE, 10GB, BS=16 | throughput (tokens/s): 7.64 -> 7.79 | Improved |
| add cache pool planning (ablation) | ZipMoE · DeepSeekV2-Lite, 10GB, BS=16 | E2E latency (s): 176.68 -> 173.23 | Improved |
| core system (lossless decompression, CPU-parallel scheduling) versus offloading baseline (ablation) | ZipMoE · DeepSeekV2-Lite, 10GB, BS=16 | throughput (tokens/s): 1.60 -> 4.43 | Improved |
| reduce OS page cache by occupying 32 GB of RAM (ablation) | ZipMoE · TTFT, Jetson | TTFT (s): 2.18 s -> 6.27 s | Hurt |
| reduce OS page cache by occupying 32 GB of RAM (ablation) | ZipMoE · TPOT, Jetson | TPOT (s): 0.46 s -> 0.75 s | Hurt |
| increase batch size so more experts are activated per layer | ZipMoE scheduler · batch inference BS 1/4/16 | throughput | Improved |
| use ZipMoE on encoder-decoder model instead of decoder-only | ZipMoE · SwitchTransformers-Large-128 | TPOT reduction: 4.99%-81.24% | Mixed |
| use off-the-shelf lossless compressors (ZSTD, LZ4HC) on MoE BF16 weights | MoE parameters · model size | compressed size: 74% (LZ4HC), 68% (ZSTD) | Improved |
Closest papers by meaning
- MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenario (NeurIPS 2026 · similarity 0.90)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.88)
- FlashMoE: Fast Distributed MoE in a Single Kernel (NeurIPS 2025 · similarity 0.87)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.87)
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.87)
- MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
