Research map / Architectures and efficiency / Attention and transformers
FutureFill: Fast Generation from Convolutional Sequence Models
Naman Agarwal, Xinyi Chen, Evan Dogariu, Devan Shah, Hubert Strauss, Vladimir Feinberg, Daniel Suo, Peter Bartlett, Elad Hazan
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 Epoched-FutureFill instead of naive online convolution for generation without prefill at 126976 tokens | FlashSTU-T STU-only · Generation time, no prefill | speedup: 1.7× speedup | Improved |
| use Epoched-FutureFill instead of naive online convolution for generation without prefill at 126976 tokens | FlashSTU-T hybrid · Generation time, no prefill | speedup: 1.5× speedup | Improved |
| use Epoched-FutureFill with a long prefill (32768) instead of naive cached convolution | FlashSTU-T STU-only · Generation time with prefill | inference time (s): 165.13 -> 75.66 | Improved |
| use Epoched-FutureFill with a short prompt (512) and total length 4096 | FlashSTU-T STU-only 670.75M · Generation time, prompt 512 | inference time (s): 12.68 -> 16.33 | Hurt |
| use Epoched-FutureFill with a short prompt (512) and total length 131072 | FlashSTU-T STU-only 670.75M · Generation time, prompt 512 | inference time (s): 1837.16 -> 1097.50 | Improved |
| set epoch length K to theoretical optimum 1024 instead of smaller K at L_gen 65536 (ablation) | Epoched-FutureFill · Generation time vs cache size K | inference time | Improved |
Closest papers by meaning
- The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation (ICLR 2025 · similarity 0.86)
- UniPrefill: Universal Long-Context Prefill Acceleration via Block-wise Dynamic Sparsification (NeurIPS 2026 · similarity 0.85)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.84)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.84)
- Orchid: Flexible and Data-Dependent Convolution for Sequence Modeling (NeurIPS 2024 · similarity 0.84)
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.84)
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.
