Research map / Architectures and efficiency / Mixture of experts
Libra: Effective yet Efficient Load Balancing for Large-scale MoE Inference
Jaehoon Yang, Yushin Kim, Seokwon Moon, Yeonhong Park, Jae W. Lee
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 |
|---|---|---|---|
| replace Lina's lookup-table predictor with lookahead gating prediction | Libra · BookCorpus (Qwen3MoE) | expert prediction accuracy (%): 47.3 -> 91.7 | Improved |
| replace Lina's lookup-table predictor with lookahead gating prediction | Libra · BookCorpus (GLM-4.5) | expert prediction accuracy (%): 11.7 -> 79.6 | Improved |
| replace Lina's lookup-table predictor with lookahead gating prediction | Libra · HellaSwag (GLM-4.5) | expert prediction accuracy (%): 11.5 -> 76.6 | Improved |
| use Libra's predictor instead of Lina's | MoE layer · MoE layer breakdown | predict overhead: 0.40 -> 0.03 | Improved |
| add Libra load balancing and overlapped execution over SGLang | Qwen3MoE MoE layer · MoE layer breakdown (seq 1024, batch 32) | total MoE layer latency: 13.61 -> 9.07 | Improved |
| use Libra instead of Lina | Qwen3MoE MoE layer · MoE layer breakdown (seq 1024, batch 32) | total MoE layer latency: 11.33 -> 9.07 | Improved |
| add Libra over SGLang at long context and large batch | Qwen3MoE · End-to-end TTFT | TTFT reduction (%): up to 27% | Improved |
| add Libra over SGLang at short sequences and small batch | Qwen3MoE · End-to-end latency, batch 8, input 1024, output 1 | latency (s): 0.36 -> 0.38 | Hurt |
| add Libra over SGLang at batch 64, input 1024 | Qwen3MoE · End-to-end latency, output 1 | latency (s): 2.72 -> 2.00 | Improved |
| add Libra over Lina at batch 64, input 2048 | Qwen3MoE · End-to-end latency, output 1 | latency (s): 4.76 -> 4.00 | Improved |
| add Libra to improve prefill throughput over state of the art | Qwen3MoE and GLM-4.5 · Prefill throughput, 4 imbalanced datasets | throughput (tokens/s): up to 19.2% | Improved |
Closest papers by meaning
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · similarity 0.89)
- Toward Efficient Inference for Mixture of Experts (NeurIPS 2024 · similarity 0.89)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.89)
- Mixture of Lookup Experts (ICML 2025 · similarity 0.89)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.88)
- FloE: On-the-Fly MoE Inference on Memory-constrained GPU (ICML 2025 · similarity 0.88)
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.
