Research map / Architectures and efficiency / Quantization
ECO: Quantized Training without Full-Precision Master Weights
Mahdi Nikdan, Amir Zandieh, Dan Alistarh, Vahab Mirrokni
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 ECO momentum error injection (SR) (ablation) | FP8 w/o master weights · C4 pretraining 30M | validation loss: 3.4008 -> 3.3317 | Improved |
| add ECO momentum error injection (SR) (ablation) | FP8 w/o master weights · C4 pretraining 430M | validation loss: 2.8378 -> 2.6544 | Improved |
| add ECO momentum error injection (SR) (ablation) | FP8 w/o master weights · C4 pretraining 800M | validation loss: 2.9471 -> 2.5399 | Improved |
| add ECO to RTN weight removal (naive RTN diverges) (ablation) | FP8 w/o master weights + RTN · C4 pretraining 30M-800M | validation loss: dvg -> 3.3640 | Improved |
| use stochastic rounding instead of RTN with ECO (ablation) | ECO · C4 pretraining 800M | validation loss: 2.6046 -> 2.5399 | Improved |
| use stochastic rounding instead of RTN with ECO (ablation) | ECO · C4 pretraining 30M | validation loss: 3.3640 -> 3.3317 | Improved |
| remove FP32 master weights with ECO (SR) versus FP8 w/ MW + SR | FP8 training · C4 pretraining 430M | validation loss: 2.6500 -> 2.6544 | No clear effect |
| remove master weights with ECO (RTN) | INT4 QAT of DeepSeek-MoE-16B · MMLU | accuracy: 37.87 -> 38.63 | No clear effect |
| remove master weights with ECO (SR) | INT4 QAT of DeepSeek-MoE-16B · GSM8K | accuracy: 16.15 -> 16.00 | No clear effect |
| remove master weights naively | INT4 QAT of DeepSeek-MoE-16B · fine-tuning training loss | training loss | Hurt |
| remove FP32 master weights to FP8 with ECO | SMoE 2.1B · peak memory | bytes per parameter: 12 -> 9 | Improved |
Closest papers by meaning
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.90)
- Treasures in Discarded Weights for LLM Quantization (AAAI 2025 · similarity 0.89)
- Scaling Laws for Precision (ICLR 2025 · similarity 0.89)
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.89)
- Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs (ICML 2024 · similarity 0.89)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 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.
