AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Blockwise Parallel Transformers for Large Context Models

Hao Liu, Pieter Abbeel

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Sparse efficient transformers for long context · 3 citations (OpenAlex)

Read the paper

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
blockwise attention and feedforward instead of MemoryEfficient attentionTransformer · 1B on 1 A100max context length: 65536 -> 131072Improved
blockwise instead of vanilla attentionTransformer · 3B on 1 A100max context length: 8192 -> 65536Improved
blockwise instead of MemoryEfficientTransformer · 70B on 64 TPUv4max context length: 2048 -> 8192Improved
blockwise instead of MemoryEfficientTransformer · 30B on 64 TPUv4max context length: 4096 -> 16384Improved
blockwise instead of MemoryEfficientTransformer · 3B at 65536 context memorymemory: 75GB -> 70GBImproved
blockwise instead of MemoryEfficientTransformer · 3B at 131072 context memorymemory: oom -> 79GBImproved
blockwise instead of vanillaTransformer · 3B at 16384 context memorymemory: oom -> 45GBImproved
blockwise instead of vanillaTransformer · GPT-XL 1B throughput, context 8192throughput: 2455 -> 2875Improved
blockwise instead of vanillaTransformer · GPT-XL 1B throughput, context 16384throughput: 1701 -> 2045Improved
blockwise instead of MemoryEfficientTransformer · GPT-XL 1B throughput, context 2048throughput: 4371 -> 3985Hurt
blockwise instead of MemoryEfficientTransformer · GPT-XL 1B throughput, context 32768throughput: 810 -> 857Improved
blockwise attention across context lengths 2048-16384Transformer · OpenWebTextvalidation loss: 2.46 -> 2.46No clear effect

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