Research map / Architectures and efficiency / RNNs and normalization
FAFO: Lossy KV Cache Compression for Lossless Inference Acceleration via Draftless Fumble Decoding
Hoang Anh Duy Le, Shaochen Zhong, Yifan Lu, Yingtong Dou, Jiayi Yuan, Yu-Neng Chuang, Xiran Fan, Guanchu Wang, Yuzhong Chen, Xia Hu
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 |
|---|---|---|---|
| parallel draft-and-verify in one forward pass instead of two forward passes (ablation) | FAFO-Stream · MT-Bench, Llama-2-7b-chat | speedup: 1.19× -> 1.91× | Improved |
| parallel verification instead of two forwards (ablation) | FAFO-Stream · MT-Bench, Llama-2-7b-chat | acceptance length: 2.18 -> 2.29 | No clear effect |
| generate guesses on compressed KV cache instead of full cache | FAFO vs Lookahead · MT-Bench, Llama-2-7b-chat | speedup: 1.61× -> 1.91× | Improved |
| increase number of n-gram guesses beyond about 10 (ablation) | Lookahead · n-gram guesses sweep | throughput | Hurt |
| increase number of n-gram guesses (ablation) | FAFO · n-gram guesses sweep | throughput | Improved |
| cache n-grams under longer suffix prefixes (lookback window) (ablation) | FAFO · n-gram guesses sweep | acceptance length | Improved |
| swap in Quest instead of StreamingLLM as compression method (ablation) | FAFO · MT-Bench, Llama-2-7b-chat | speedup: 1.91× -> 1.32× | Hurt |
| use FAFO instead of TriForce | FAFO-Stream · MT-Bench, Llama-2-7b-chat | speedup: 0.21× -> 1.91× | Improved |
| use FAFO instead of Lookahead on long-context LongBench | FAFO-Stream · LongBench Qasper, Llama-3.1-8B | speedup: 0.85× -> 2.20× | Improved |
| use FAFO-Stream vs Lookahead on Qwen2.5-32B GSM8K | FAFO-Stream · GSM8K, Qwen2.5-32B-Instruct | speedup: 1.30× vs 1.44× Lookahead | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SWIFT — SwiFT: Swin 4D fMRI Transformer (NeurIPS 2023) · better on 6 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CompilerKV: Risk-Adaptive KV Cache Compression via Offline Experience Compilation (NeurIPS 2026 · similarity 0.90)
- Get More with LESS: Synthesizing Recurrence with KV Cache Compression for Efficient LLM Inference (ICML 2024 · similarity 0.89)
- Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction (NeurIPS 2026 · similarity 0.89)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.89)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.88)
- MUSTAFAR: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference (NeurIPS 2025 · similarity 0.88)
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.
