AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

AutoSP: Unlocking Long-Context LLM Training Via Compiler-Based Sequence Parallelism

Ahan Gupta, Zhihao Wang, Neel Dani, Masahiro Tanaka, Olatunji Ruwase, Minjia Zhang

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add SP-aware AC pass on top of SP pass (ablation)AutoSP · Max trainable tokensmax tokens before OOM: 77,000 -> 128,000Improved
add SP-aware AC pass on top of SP pass (ablation)AutoSP · 40k sequence lengthiteration time (s): 1.09 -> 1.19Hurt
replace hand-written DS-Ulysses with AutoSP SP pass only (ablation)Llama-3.1 1B · Max trainable tokensmax tokens before OOM: 81,000 -> 77,000No clear effect
replace hand-written DS-Ulysses with AutoSP SP pass only (ablation)Llama-3.1 1B · 40k sequence lengthiteration time (s): 1.06 -> 1.09Hurt
replace DS-Ulysses with full AutoSP (SP + AC)Llama-3.1 1B · Max trainable tokensmax tokens before OOM: 81,000 -> 128,000 (1.58x)Improved
use AutoSP instead of ZeRO-3 (FSDP) with torch.compileLlama 3B/8B/13B · Max sequence length on 8x A100max sequence length: upto 5x, 5.6x and 2.5x longerImproved
use AutoSP instead of inductor-compiled DS-UlyssesLlama 3B/8B/13B · Max sequence length on 8x A100max sequence length: upto 2.14x, 3x and 1.88xImproved
use AutoSP instead of RingAttentionLlama 3B/8B/13B · Max sequence length on 8x A100max sequence length: upto 2.14x, 3x and 1.6xImproved
use AutoSP instead of ZeRO-3 compiled baselineLlama-3.1 8B · Per-iteration time, 8 GPUsiteration time: 5x lower per-iteration timeImproved
use AutoSP instead of inductor baseline on NVIDIA and AMDLlama-3.2 1B/3B · Runtime on GH200/MI250speedup: 0.97x, 0.98x (NVIDIA); 0.97x, 0.87x (AMD)Mixed
use AutoSP instead of inductor baseline on diverse hardwareLlama-3.2 1B/3B · Max sequence lengthmax sequence length: 1.58x (2.70x) NVIDIA; 2x (2.5x) AMDImproved
add AutoSP AC-pass rematerializing matmuls (ablation)Llama-3.2 1B at 40k · Operator activation memorymemory reduction: 13.03x (attention), 2.22x (MLP)Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy