Research map / Architectures and efficiency / Attention and transformers
Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention
Haiquan Qiu, Quanming Yao
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 |
|---|---|---|---|
| disable tiling by setting block size equal to sequence length (ablation) | flash attention 2 · GPT-2S BF16 OpenWebText | training stability | No clear effect |
| replace flash attention with standard attention in layer 2 (ablation) | GPT-2 flash attention BF16 · GPT-2S BF16 OpenWebText | training stability | Improved |
| use flash attention only in layer 2 (ablation) | GPT-2 flash attention BF16 · GPT-2S BF16 OpenWebText | training stability | Hurt |
| compute delta as rowsum(dP*P) instead of rowsum(dO*O) (ablation) | flash attention backward · GPT-2S BF16 OpenWebText | training stability | Improved |
| recompute O as PV in FP32 for delta in backward (ablation) | flash attention backward · GPT-2S BF16 OpenWebText | training stability | Improved |
| compute O in FP32 in forward pass (ablation) | flash attention forward · GPT-2S BF16 OpenWebText | training stability | Improved |
| compute O in high precision only for outlier heads 1,7,8,9,11,12 (ablation) | flash attention · GPT-2S BF16 OpenWebText | training stability | Improved |
| modify softmax normalization to avoid P entries exactly equal to 1 (beta=2) | flash attention · GPT-2S with AdamW and Muon, GPT-2M AdamW | loss explosion | Improved |
| use standard BF16 flash attention | GPT-2 · GPT-2S BF16 OpenWebText | validation loss | Hurt |
Closest papers by meaning
- Small-scale proxies for large-scale Transformer training instabilities (ICLR 2024 · similarity 0.89)
- Transformers Learn Low Sensitivity Functions: Investigations and Implications (ICLR 2025 · similarity 0.87)
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.86)
- Unraveling the Gradient Descent Dynamics of Transformers (NeurIPS 2024 · similarity 0.86)
- Linear attention is (maybe) all you need (to understand Transformer optimization) (ICLR 2024 · similarity 0.85)
- What Happens During the Loss Plateau? Understanding Abrupt Learning in Transformers (NeurIPS 2025 · 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.
