Research map / Architectures and efficiency / Attention and transformers
Blockwise Parallel Transformers for Large Context Models
Hao Liu, Pieter Abbeel
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 |
|---|---|---|---|
| blockwise attention and feedforward instead of MemoryEfficient attention | Transformer · 1B on 1 A100 | max context length: 65536 -> 131072 | Improved |
| blockwise instead of vanilla attention | Transformer · 3B on 1 A100 | max context length: 8192 -> 65536 | Improved |
| blockwise instead of MemoryEfficient | Transformer · 70B on 64 TPUv4 | max context length: 2048 -> 8192 | Improved |
| blockwise instead of MemoryEfficient | Transformer · 30B on 64 TPUv4 | max context length: 4096 -> 16384 | Improved |
| blockwise instead of MemoryEfficient | Transformer · 3B at 65536 context memory | memory: 75GB -> 70GB | Improved |
| blockwise instead of MemoryEfficient | Transformer · 3B at 131072 context memory | memory: oom -> 79GB | Improved |
| blockwise instead of vanilla | Transformer · 3B at 16384 context memory | memory: oom -> 45GB | Improved |
| blockwise instead of vanilla | Transformer · GPT-XL 1B throughput, context 8192 | throughput: 2455 -> 2875 | Improved |
| blockwise instead of vanilla | Transformer · GPT-XL 1B throughput, context 16384 | throughput: 1701 -> 2045 | Improved |
| blockwise instead of MemoryEfficient | Transformer · GPT-XL 1B throughput, context 2048 | throughput: 4371 -> 3985 | Hurt |
| blockwise instead of MemoryEfficient | Transformer · GPT-XL 1B throughput, context 32768 | throughput: 810 -> 857 | Improved |
| blockwise attention across context lengths 2048-16384 | Transformer · OpenWebText | validation loss: 2.46 -> 2.46 | No clear effect |
Closest papers by meaning
- RingAttention with Blockwise Transformers for Near-Infinite Context (ICLR 2024 · similarity 0.91)
- The Sharpness Disparity Principle in Transformers for Accelerating Language Model Pre-Training (ICML 2025 · similarity 0.88)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.88)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.87)
- An Evolved Universal Transformer Memory (ICLR 2025 · similarity 0.87)
- Parallelizing Linear Transformers with the Delta Rule over Sequence Length (NeurIPS 2024 · similarity 0.87)
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.
