Research map / Architectures and efficiency / Quantization
Pushing the Limits of Block Rotations in Post-Training Quantization
Sai Sanjeet, Ian Colbert, Pablo Monteagudo-Lago, Giuseppe Franco, Yaman Umuroglu, Nicholas J. Fraser
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 |
|---|---|---|---|
| add MassDiff permutation before block rotation (ablation) | Llama3 1B block 16 Qronos INT4 · WikiText2 | WikiText2 perplexity: 35.8 -> 18.2 | Improved |
| add MassDiff permutation before block rotation (ablation) | Llama3 8B block 32 Qronos INT4 · WikiText2 | WikiText2 perplexity: 12.8 -> 8.5 | Improved |
| add MassDiff permutation before block rotation (ablation) | Llama3 1B block 32 RTN INT4 · WikiText2 | WikiText2 perplexity: 4e3 -> 38.0 | Improved |
| add MassDiff permutation before block rotation (ablation) | Llama3 8B block 32 RTN INT4 · WikiText2 | WikiText2 perplexity: 6e3 -> 12.0 | Improved |
| add MassDiff permutation to MR-Qronos on GSM8K (ablation) | Llama3 8B INT4 · GSM8K | accuracy: 1.7 -> 75.4 | Improved |
| use MassDiff instead of random permutation (ablation) | PeRQ* Llama3 1B · WikiText2 | WikiText2 perplexity: 27.8 -> 16.9 | Improved |
| use MassDiff instead of absmax permutation (ablation) | PeRQ* Llama3 1B · WikiText2 | WikiText2 perplexity: 20.1 -> 16.9 | Improved |
| use MassDiff instead of ZigZag permutation (ablation) | PeRQ* Llama3 1B · WikiText2 | WikiText2 perplexity: 17.1 -> 16.9 | Improved |
| use MassDiff instead of ZigZag with 128K calibration tokens (ablation) | PeRQ* Llama3 1B block 32 · WikiText2 | WikiText2 perplexity: 17.3 -> 16.9 | Improved |
| add MassDiff with C4 calibration (ablation) | PeRQ* Llama3 1B · WikiText2 | WikiText2 perplexity: 29.2 -> 18.9 | Improved |
| replace GPTQ with Qronos rounding (ablation) | MassDiff + QuaRot Llama3 1B · WikiText2 | WikiText2 perplexity: 28.4 -> 16.9 | Improved |
| replace RTN with GPTQ rounding when learning rotations (ablation) | MassDiff + SpinQuant Llama3 1B · WikiText2 | WikiText2 perplexity: 15.9 -> 19.5 | Hurt |
Closest papers by meaning
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.89)
- QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024 · similarity 0.87)
- DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · similarity 0.87)
- Block Sphere Vector Quantization (NeurIPS 2026 · similarity 0.87)
- DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization (NeurIPS 2025 · similarity 0.87)
- KroQuant: Kronecker-Structured Block Transforms for Efficient Post-Training Quantization of Diffusion Transformers (NeurIPS 2026 · 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.
