Research map / Architectures and efficiency / Quantization
STaMP: Sequence Transformation and Mixed Precision for Low-Precision Activation Quantization
Marco Federici, Riccardo Del Chiaro, Boris van Breugel, Paul N. Whatmough, Markus Nagel
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 STaMP sequence transform with 64 8-bit tokens | FlatQuant · WikiText-2 W4A4KV4 | perplexity: Llama 3 8B: 6.89 -> 6.77 | Improved |
| add STaMP sequence transform | QuaRot · WikiText-2 W4A4KV4, Qwen 2.5 3B it | perplexity: 94.86 -> 71.13 | Improved |
| add STaMP sequence transform | RTN · WikiText-2 W4A4KV4, Llama 3 8B | perplexity: 668 -> 95.3 | Improved |
| add STaMP to RTN | RTN on PixArt-Sigma · COCO W4A4 | Image Reward: 0.38 -> 0.80 | Improved |
| add STaMP to SVDQuant | SVDQuant on PixArt-Sigma · COCO W4A4 | Image SQNR: 8.78 -> 9.72 | Improved |
| add STaMP to ViDiT-Q | ViDiT-Q on PixArt-Sigma · COCO W4A4 | Image SQNR: 7.82 -> 6.37 | Hurt |
| add STaMP to ViDiT-Q | ViDiT-Q on SANA · COCO W4A4 | Image Reward: 0.89 -> 0.86 | Hurt |
| add STaMP to QuaRot feature transform (ablation) | QuaRot on PixArt-Sigma ffn.down_proj · A4 single-activation SQNR | Image SQNR: 8.42 -> 14.91 | Improved |
| apply STaMP to attn2.to_out activations (ablation) | PixArt-Sigma · A4 single-activation SQNR | Image SQNR: 15.48 -> 15.49 | No clear effect |
| use DWT instead of Hadamard for the sequence transform (ablation) | STaMP · PixArt-Sigma denoising step | CUDA latency overhead (%): 56.8 -> 4.8 | Improved |
| use DCT, WHT or DWT as sequence transform (ablation) | STaMP · PixArt-Sigma and Llama v3 8B, A4 | quality | No clear effect |
| increase number of 8-bit high-precision tokens (ablation) | STaMP · PixArt-Sigma, activation-only quantization | SQNR | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- FlatQuant — FlatQuant: Flatness Matters for LLM Quantization (ICML 2025) · better on 4 of 4 shared results
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 3 of 3 shared results
- SmoothQuant — SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models (ICML 2023) · better on 1 of 1 shared result
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
- ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs (NeurIPS 2026 · similarity 0.88)
- ASER: Activation Smoothing and Error Reconstruction for Large Language Model Quantization (AAAI 2025 · similarity 0.86)
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.86)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.86)
- MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods (ICLR 2025 · similarity 0.85)
- LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models (ICLR 2024 · similarity 0.85)
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.
