Reproduce “AutoSP: Unlocking Long-Context LLM Training Via Compiler-Based Sequence Parallelism” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ahan Gupta, Zhihao Wang, Neel Dani, Masahiro Tanaka, Olatunji Ruwase, Minjia Zhang
Researchers at the University of Illinois Urbana-Champaign, Anyscale, and Snowflake developed AutoSP, a compiler-based, PyTorch-native framework that automates sequence parallelism and introduces a sequence-parallel aware activation checkpointing pass for long-context Large Language Model (LLM) training. AutoSP enables up to 3x longer input sequences and up to 5x faster training iteration times compared to existing methods, while maintaining near-baseline runtime performance on NVIDIA and AMD GPUs.
Key insights
- Torch-IR's higher abstraction level is more suitable than Aten-IR for identifying computational graph regions, simplifying tensor resizing and collective insertion for sequence parallelism.
- The relative computational cost of linear projections and MLP layers decreases significantly as the input sequence length increases, enabling more aggressive activation rematerialization for memory savings.
- Automating distributed training optimizations, including sequence parallelism and memory management, within a deep-learning compiler improves developer productivity and portability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.