Research map / Architectures and efficiency / Quantization
TurboBoA: Faster and Exact Attention-aware Quantization without Backpropagation
Junhan Kim, Yeo Jeong Park, Seungwoo Son, Chungman Lee, Ho-young Kim, Joonyoung Kim, Yongkweon Jeon
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 |
|---|---|---|---|
| quantize N=16 out-channels jointly instead of one by one (F1) (ablation) | BoA · Quantization time, INT2 Llama3-8B | time (min): 94.75 -> 25.30 | Improved |
| quantize N=16 out-channels jointly (F1) (ablation) | BoA · Wiki2 PPL, INT2 Llama3-8B | perplexity: 15.20 -> 15.41 | No clear effect |
| quantize N=64 jointly instead of N=16 (ablation) | TurboBoA F1 · Quantization time Llama3-8B | time (min): 25.30 -> 21.56 | No clear effect |
| add error compensation for pre-quantized layers (F2) (ablation) | TurboBoA · Wiki2 PPL INT2 Llama3.2-1B | perplexity: 41.85 -> 37.15 | Improved |
| add adaptive grid with CD refinement (F3) (ablation) | TurboBoA · Wiki2 PPL INT2 Llama3.2-1B | perplexity: 41.85 -> 39.45 | Improved |
| add both F2 and F3 (ablation) | TurboBoA · Wiki2 PPL INT2 Llama3.2-1B | perplexity: 41.85 -> 33.33 | Improved |
| add F3 (adaptive grid) to TurboBoA on Llama3.2-1B C4 (ablation) | TurboBoA · C4 PPL INT2 Llama3.2-1B | perplexity: 108.1 -> 107.3 | Improved |
| add error compensation F2 (extra FP forward pass) (ablation) | TurboBoA · quantization time, Llama3-8B | time (min): 25.30 -> 40.16 | Hurt |
| replace BoA with TurboBoA in weight-activation quantization | SpinQuant W2A4KV4 · Wiki2 PPL | perplexity: 77.05 -> 63.07 (Llama3.2-1B) | Improved |
| replace BoA with TurboBoA for 3-bit weight-only | QuaRot INT3 · Wiki2 PPL Llama3-8B | perplexity: 7.145 -> 7.116 | No clear effect |
Closest papers by meaning
- BoA: Attention-aware Post-training Quantization without Backpropagation (ICML 2025 · similarity 0.92)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 2024 · similarity 0.89)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.89)
- GANQ: GPU-Adaptive Non-Uniform Quantization for Large Language Models (ICML 2025 · similarity 0.88)
- Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization (NeurIPS 2025 · similarity 0.88)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · 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.
