Research map / Architectures and efficiency / Quantization
Attn-QAT: 4-Bit Attention With Quantization-Aware Training
Peiyuan Zhang, Matthew Noto, Wenxuan Tan, Chengquan Jiang, Will Lin, Wei Zhou, Hao Zhang
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 high-precision output O' for the backward pass (ablation) | Attn-QAT · VBench (Wan 2.1 1.3B) | Overall Quality: 0.7185 -> 0.8252 | Improved |
| add fake quantization of P in backward recomputation (ablation) | Attn-QAT · VBench (Wan 2.1 1.3B) | Overall Quality: 0.8254 -> 0.8252 | No clear effect |
| add K smoothing to Attn-QAT (ablation) | Attn-QAT · VBench (Wan 2.1 1.3B) | Overall Quality: 0.8252 -> 0.8232 | No clear effect |
| add two-level quantization of P to Attn-QAT (ablation) | Attn-QAT · VBench (Wan 2.1 1.3B) | Overall Quality: 0.8252 -> 0.8257 | No clear effect |
| apply Attn-QAT training to FP4 attention instead of training-free FP4 | Wan 2.1 1.3B · VBench (Wan 2.1 1.3B) | Overall Quality: 0.7785 -> 0.8252 | Improved |
| apply Attn-QAT training instead of SageAttention3 | Wan 2.1 14B · VBench (Wan 2.1 14B) | Overall Quality: 0.8203 -> 0.8279 | Improved |
| apply Attn-QAT instead of SageAttention3 on Wan 1.3B | Wan 2.1 1.3B · VBench (Wan 2.1 1.3B) | Imaging Quality: 0.5507 -> 0.6775 | Improved |
| add Attn-QAT to FP4 attention (continued pretraining) | Qwen3-14B · WinoGrande | accuracy: 0.7214 -> 0.7585 | Improved |
| add Attn-QAT to FP4 attention (continued pretraining) | Llama 3.1-70B · MMLU | accuracy: 0.7577 -> 0.7773 | Improved |
| add Attn-QAT to FP4 attention (continued pretraining) | Llama 3.1-70B · WikiText | perplexity: 0.3275 -> 0.3076 | Improved |
| use Attn-QAT instead of BF16 attention in SFT | Qwen3-14B · MMLU-Redux (SFT) | accuracy: 0.8316 -> 0.8392 | No clear effect |
| use Attn-QAT instead of BF16 attention in SFT | Llama 3.1-70B · GSM8K (SFT) | accuracy: 0.8840 -> 0.8673 | Hurt |
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.
- SageAttention3 — SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training (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
- SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training (NeurIPS 2025 · similarity 0.88)
- Compute-Optimal Quantization-Aware Training (ICLR 2026 · similarity 0.87)
- SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization (ICML 2025 · similarity 0.86)
- FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision (NeurIPS 2024 · similarity 0.84)
- SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration (ICLR 2025 · similarity 0.84)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.84)
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.
