Research map / Architectures and efficiency / RNNs and normalization
AMP: Automatically Finding Model Parallel Strategies with Heterogeneity Awareness
Dacheng Li, Hongyi Wang, Eric Xing, Hao Zhang
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 |
|---|---|---|---|
| use AMP instead of Megatron heuristics | GPT-2 · heterogeneous cluster | min iteration time: 1.97 -> 1.28 s (1.54x speedup) | Improved |
| use AMP instead of Megatron heuristics | TransGAN generator · heterogeneous model | min iteration time: 1.89 -> 1.07 s (1.77x) | Improved |
| use AMP instead of Megatron on homogeneous setup | GPT-2 · homogeneous cluster | min iteration time: 1.32 -> 1.20 s | Improved |
| use DP-based layer assignment instead of uniform assignment (ablation) | AMP best strategy · heterogeneous model | min iteration time: 1.25 -> 1.07 s (1.17x) | Improved |
| use DP-based layer assignment instead of parameter-balanced assignment (ablation) | AMP best strategy · heterogeneous model | min iteration time: 2.19 -> 1.07 s | Improved |
| add simulated annealing device placement search (ablation) | AMP · homogeneous cluster | min iteration time: 1.20 -> 1.57 s | Hurt |
| add simulated annealing device placement search (ablation) | AMP · heterogeneous cluster | min iteration time: 1.28 -> 1.46 s | Hurt |
| cost-model guided search instead of random strategy selection (ablation) | AMP · heterogeneous cluster | mean iteration time: 3.04 -> 1.73 s | Improved |
Closest papers by meaning
- Pfeife: Automatic Pipeline Parallelism for PyTorch (ICML 2025 · similarity 0.86)
- Synergistic Tensor and Pipeline Parallelism (NeurIPS 2025 · similarity 0.84)
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.84)
- UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming (CVPR 2025 · similarity 0.84)
- Towards a Better Theoretical Understanding of Independent Subnetwork Training (ICML 2024 · similarity 0.83)
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.83)
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.
