Research map / Architectures and efficiency / Quantization
TetraJet-v2: Accurate NVFP4 Training for Large Language Models with Oscillation Suppression and Outlier Control
Yuxiang Chen, Yifan Liu, Xiaoming Xu, Pengle Zhang, Michael Beyer, Martin Rapp, Jun Zhu, Jianfei Chen
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 OsciReset (reset oscillating master weights to bin centers) (ablation) | TetraJet-v2-base+OC · OLMo2-150M, 107B tokens | validation perplexity | Improved |
| add Q-EMA, Freeze or Dampen oscillation suppression instead of OsciReset (ablation) | TetraJet-v2-base+OC · OLMo2-150M, 107B tokens | validation perplexity | Hurt |
| add OutControl (retain top 10% activation outlier channels in higher precision) (ablation) | TetraJet-v2-base · OLMo2-370M, 212B tokens | validation perplexity | Improved |
| add OsciReset and OutControl to TetraJet-v2-base (ablation) | TetraJet-v2-base · OLMo2-370M, C4 validation | validation perplexity: 25.50 -> 25.11 | Improved |
| add OsciReset and OutControl to TetraJet-v2-base (ablation) | TetraJet-v2-base · OLMo2-70M, C4 validation | validation perplexity: 49.33 -> 47.75 | Improved |
| apply Random Hadamard Transform in the forward pass (ablation) | NVFP4 linear layer · OLMo2 NVFP4 training | optimization quality | Hurt |
| apply RHT to both dX and dW backward matmuls (ablation) | NVFP4 linear layer · OLMo2 NVFP4 training | optimization quality | Improved |
| quantize both backward matmuls rather than one (ablation) | NVFP4 linear layer · OLMo2-370M, 52B tokens | validation perplexity delta vs BF16: -0.01 and +0.07 individually -> +0.21 combined | Hurt |
| keep mlp.ffn2 in BF16 (ablation) | fully quantized NVFP4 model · OLMo2-370M, 52B tokens | validation perplexity: 24.89 -> 24.54 | Improved |
| switch to FP6 x FP4 for the last 50B tokens | TetraJet-v2-full · OLMo2-370M, 212B tokens | validation perplexity | Improved |
| use stochastic rounding, aligned activations and 1x16 weight blocks instead of NVIDIA recipe design | NVFP4 linear layer · OLMo2-370M, 212B tokens | validation perplexity: 26.20 -> 25.50 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Quartet — Quartet: Native FP4 Training Can Be Optimal for Large Language Models (NeurIPS 2025) · better on 4 of 4 shared results
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
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.91)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.88)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.88)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.88)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.88)
- Quartet: Native FP4 Training Can Be Optimal for Large Language Models (NeurIPS 2025 · 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.
