Research map / Architectures and efficiency / Quantization
Metis: Training LLMs with FP4 Quantization
Hengjie Cao, Mengyi Chen, Yifeng Yang, Fang Dong, Ruijun Huang, Jixian Zhou, Anrui Chen, Mingzhi Dong, Yujiang Wang, Jinlong Hou, Yuan Cheng, FAN WU and 4 more
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 |
|---|---|---|---|
| Apply Metis spectral decomposition instead of direct NVFP4 | LLaMA-3 8B FP4 training · DCLM 100B tokens | test loss: 2.53 -> 2.45 | Improved |
| Apply Metis spectral decomposition instead of direct NVFP4 | GPT-2 130M FP4 training · DCLM 100B tokens | test loss: 3.32 -> 3.20 | Improved |
| Apply Metis spectral decomposition instead of direct NVFP4 | GPT-2 1.1B FP4 training · DCLM 100B tokens | test loss: 3.22 -> 3.01 | Improved |
| Apply Metis instead of direct NVFP4 | LLaMA-3 8B FP4 training · Downstream average (7 tasks) | average accuracy: 50.0 -> 51.0 | Improved |
| Apply Metis instead of direct NVFP4 | GPT-2 1.1B FP4 training · Downstream average (7 tasks) | average accuracy: 47.5 -> 48.6 | Improved |
| Add spectral decomposition of gradients (ablation) | Metis on LLaMA-3 8B · DCLM 100B tokens | test loss: 2.51 -> 2.45 | Improved |
| Add spectral decomposition of gradients (ablation) | Metis on LLaMA-3 8B · Downstream average | average accuracy: 50.1 -> 51.1 | Improved |
| Add spectral decomposition of weights (ablation) | Metis on LLaMA-3 8B · DCLM 100B tokens | test loss: 2.47 -> 2.45 | Improved |
| Add spectral decomposition of activations (ablation) | Metis on LLaMA-3 8B · DCLM 100B tokens | test loss: 2.46 -> 2.45 | Improved |
| Add spectral decomposition of weights (ablation) | Metis on LLaMA-3 8B · Downstream average | average accuracy: 50.7 -> 51.1 | Improved |
| Add spectral decomposition of activations (ablation) | Metis on LLaMA-3 8B · Downstream average | average accuracy: 50.9 -> 51.1 | Improved |
| Use sparse random sampling instead of full-batch decomposition (ablation) | Metis on LLaMA-3 8B · DCLM 100B tokens | test loss: 2.44 -> 2.45 | No clear effect |
Closest papers by meaning
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.89)
- any4: Learned 4-bit Numeric Representation for LLMs (ICML 2025 · similarity 0.88)
- Quartet: Native FP4 Training Can Be Optimal for Large Language Models (NeurIPS 2025 · similarity 0.87)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.87)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.87)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · 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.
