Research map / Architectures and efficiency / RNNs and normalization
JanusPipe: Efficient Pipeline Parallel Training for Machine Learning Interatomic Potentials
Hongyu Wang, Weijian Liu, Hongtao Xu, Yan Wang, Mingzhen Li, Weile Jia, Guangming Tan
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 |
|---|---|---|---|
| co-locate FE and FF on same device and double pipeline stages (SymFold) (ablation) | 1F1B-2nd · UMA/eSEN training, 32 GPUs | throughput: up to 23% | Improved |
| add WaveK schedule on top of SymFold (ablation) | JanusPipe · UMA/eSEN training, 32 GPUs | throughput: further 18% | Improved |
| enable GARS micro-batch repacking (ablation) | JanusPipe · UMA/eSEN training | throughput: ~11% on average | Improved |
| tune WaveK unit size k offline (k=Best) instead of k=P (ablation) | JanusPipe · UMA-2.3B (P=8, G=2, D=2) | relative throughput: 1.55x -> 1.64x | Improved |
| use finer-grained wave-style Hanayo schedule instead of 1F1B for four-phase MLIPs | baseline pipeline · UMA-1.2B (P=8, G=2, D=2) | throughput: 1.00x -> 1.03x | No clear effect |
| use Hanayo wave schedule instead of 1F1B | baseline pipeline · UMA models | peak GPU memory: JanusPipe k=P saves up to 42.70% vs Hanayo-2nd | Hurt |
| use JanusPipe (k=P) instead of 1F1B-2nd | conservative MLIP pipeline · UMA/eSEN | peak GPU memory: up to 20.56% reduction | Improved |
| increase WaveK unit size k from 4 to 8 (ablation) | JanusPipe · UMA-1.2B (P=4, G=D=1) | speedup: 1.05x over k=4 | Improved |
| increase pipeline degree P from 1 to 4 (ablation) | JanusPipe · UMA-2.3B (G=4) | peak GPU memory: 38.74 GB -> 12.84 GB | Improved |
| use SymFold finer partition instead of baselines | eSEN-220M · eSEN-220M (P=8, G=2, D=2) | feasibility (OOM) | Improved |
Closest papers by meaning
- Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · similarity 0.88)
- The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.88)
- FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials (ICML 2025 · similarity 0.87)
- Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials (NeurIPS 2026 · similarity 0.87)
- MLIP Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials via an Open, Accessible Benchmark Platform (NeurIPS 2025 · similarity 0.86)
- Colour me shocked: Exact Molecular Hessians from MLIPs in O(N) time using sparse differentiation! (NeurIPS 2026 · 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.
