Research map / Architectures and efficiency / Mixture of experts
Least-Loaded Expert Parallelism: Load Balancing An Imbalanced Mixture-of-Experts
Xuan-Phi Nguyen, Shrey Pandit, Austin Xu, Caiming Xiong, Shafiq Joty
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 |
|---|---|---|---|
| reroute excess tokens and expert weights to least-loaded GPUs (LLEP) instead of standard EP | LLEP vs EP · MoE layer, extreme imbalance (95% into 1 expert) | speedup: up to 6.11x (gpt-oss-120b layer) | Improved |
| use LLEP instead of EP on peak memory per GPU | LLEP vs EP · MoE layer under imbalance | peak memory: EP up to 4x growth; LLEP near constant | Improved |
| use LLEP for full-model forward throughput | LLEP vs EP · gpt-oss-20b / gpt-oss-120b on Nemotron-Math | throughput: 2.2× / 1.88× | Improved |
| use LLEP for full-parameter fine-tuning | LLEP vs EP · gpt-oss-20b AIME'25 | wall-time to accuracy: 1.25× speedup, comparable accuracy | Improved |
| use LLEP in pretraining DeepSeek-V3-scale model with bias load balancing | LLEP vs EP · 600B pretraining from scratch | speedup: ~1.55× | Improved |
| use LLEP instead of EPLB for prefill | LLEP vs EPLB · gpt-oss-120b prefill, 32 GPUs | throughput over EP: 1.45× -> 1.94× | Improved |
| add adaptive ratio lambda fallback to standard EP (ablation) | LLEP · balanced routing | speed | Improved |
| increase batch size (ablation) | LLEP · speedup vs EP | speedup | Improved |
| increase capacity factor alpha (ablation) | LLEP · speedup vs EP | speedup | Hurt |
| use higher lambda at low imbalance (15-20%) with small batch (ablation) | LLEP · speedup vs EP, B=8K | speedup | Improved |
| increase hidden size D, H (ablation) | LLEP · speedup vs EP | speedup | Improved |
| apply LLEP in decode with small per-GPU batch | LLEP · gpt-oss-120b decode | speedup: degrades to standard EP | No clear effect |
Closest papers by meaning
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.89)
- Shortcut-connected Expert Parallelism for Accelerating Mixture of Experts (ICML 2025 · similarity 0.89)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.88)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · similarity 0.88)
- Occult: Optimizing Collaborative Communications across Experts for Accelerated Parallel MoE Training and Inference (ICML 2025 · similarity 0.88)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.87)
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.
