Research map / Architectures and efficiency / Mixture of experts
CommitMoE: Efficient Fallback-Free MoE Inference with Offloading Under GPU Memory Constraints
Han Li, Jingwei Sun, Junqing Lin, Guangzhong Sun
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 mispredicted experts with predicted ones when no predicted expert is correct (ablation) | Qwen1.5-MoE-Chat · GSM8K | accuracy: 53.22% -> 54.51% | No clear effect |
| replace mispredicted experts with predicted ones when no predicted expert is correct (ablation) | DeepSeek-V2-Lite-Chat · BBH | accuracy: 49.12% -> 49.16% | No clear effect |
| add Output Weight Adjustment with tuned alpha1, alpha2 (ablation) | CommitMoE · DeepSeek-V2-Lite-Chat average | average score: 48.16 -> 48.48 | Improved |
| add Output Weight Adjustment with tuned alpha1, alpha2 (ablation) | CommitMoE · Qwen1.5-MoE-Chat average | average score: 42.38 -> 43.06 | Improved |
| add Output Weight Adjustment with tuned alpha1, alpha2 (ablation) | CommitMoE · Mixtral-8x7B-Instruct average | average score: 51.02 -> 50.73 | Hurt |
| use tuned model-specific OWA hyperparameters instead of neutral alpha1 = alpha2 = 1.0 (ablation) | CommitMoE OWA · Qwen1.5-MoE-Chat average | average score: 42.52 -> 43.06 | Improved |
| use tuned model-specific OWA hyperparameters instead of neutral alpha1 = alpha2 = 1.0 (ablation) | CommitMoE OWA · Mixtral-8x7B-Instruct average | average score: 50.45 -> 50.73 | Improved |
| fallback-free CommitMoE versus prefetching with fallback (+Prefetch using Commit Router) (ablation) | CommitMoE · end-to-end latency, Environment 1 | speedup | Improved |
| CommitMoE versus vLLM on slower GPU and PCIe | CommitMoE · Environment 2 (RTX 2080Ti, PCIe Gen3) | speedup: up to 4.3x Mixtral, 9.4x Qwen, 7.5x DeepSeek | Improved |
| CommitMoE versus vLLM on faster hardware | CommitMoE · Environment 1 (RTX 4090) | speedup: 1.3-1.5x Mixtral; up to 1.8x Qwen/DeepSeek | Improved |
Closest papers by meaning
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.90)
- FlashMoE: Fast Distributed MoE in a Single Kernel (NeurIPS 2025 · similarity 0.90)
- Fiddler: CPU-GPU Orchestration for Fast Inference of Mixture-of-Experts Models (ICLR 2025 · similarity 0.89)
- Toward Efficient Inference for Mixture of Experts (NeurIPS 2024 · similarity 0.89)
- MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems (NeurIPS 2025 · similarity 0.89)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 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.
