Research map / Architectures and efficiency / Mixture of experts
SMIDT: High-Performance Inference Framework for MoE Models with Dynamic Top-K Routing
Zewen Jin, Shen Fu, Chengjie Tang, Youhui Bai, Shengnan Wang, Jiaan Zhu, Chizheng Fang, Ping Gong, Cheng Li
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 |
|---|---|---|---|
| add module-level fine-grained pipeline partitioning with search (ablation) | SMIDT · PaulGrahamEssays | throughput speedup: +10% | Improved |
| add module-level fine-grained pipeline partitioning with search (ablation) | SMIDT · MMLU | throughput speedup: +6% | Improved |
| add DPMoE expert replication using spare memory (ablation) | SMIDT · PaulGrahamEssays and MMLU | throughput speedup: +4% and +14% | Improved |
| add asynchronous PP communication (ablation) | SMIDT · PaulGrahamEssays and MMLU | throughput speedup: +2% and +10% | Improved |
| use dynamic Top-K-aware search vs hand-tuned PP | hand-tuned PP baseline · prefill-only tasks (8 A40) | throughput: 1.17-1.31x | Improved |
| apply SMIDT on NVLink A100 | SGLang baseline · MMLU, DeepSeek-V2 and Qwen3-235B-A22B | throughput: 1.14-1.31x and 1.18-1.37x | Improved |
| apply SMIDT under PD disaggregation | PP-8 baseline · ShareGPT | TTFT: 1.10-1.17x mean, 1.21-1.26x P99 | Improved |
| higher PP degree beyond 8 | baselines · Qwen3-235B-A22B on 32 A40 | throughput | Hurt |
| summarization with decoding vs prefill-only | SMIDT · PD colocation | throughput gain: 1.71-1.89x vs 1.78-3.08x | Mixed |
Closest papers by meaning
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.90)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.90)
- Unchosen Experts Can Contribute Too: Unleashing MoE Models’ Power by Self-Contrast (NeurIPS 2024 · similarity 0.90)
- Mixture of Lookup Experts (ICML 2025 · similarity 0.90)
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.90)
- Shortcut-connected Expert Parallelism for Accelerating Mixture of Experts (ICML 2025 · similarity 0.90)
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.
