Approximating Softmax in Pretrained LLMs: Model Sensitivity and Kernel Acceleration

arXiv:2609.33586 · cs.LG · Submitted 2026-09-27 · Read on arXiv

cs.LG

Submitted: 2026-09-27

Updated: 2026-09-27

Terminology

• Quantization — 31× in this paper · explained in 12 episodes
• Perplexity — 28× in this paper · explained in 3 episodes
• Fidelity — 26× in this paper · explained in 2 episodes
• Transformer — 13× in this paper · explained in 2 episodes
• Core — 12× in this paper · explained in 2 episodes
• Accuracy — 9× in this paper · explained in 1 episode
• Calibration — 5× in this paper · explained in 7 episodes
• Robustness — 5× in this paper · explained in 5 episodes
• Frozen model — 4× in this paper · explained in 3 episodes
• Trace — 3× in this paper · explained in 2 episodes
• Ablation — 2× in this paper · explained in 1 episode
• Benchmark — 2× in this paper · explained in 2 episodes
• Decode — 2× in this paper · explained in 2 episodes
• Generative AI — 2× in this paper · explained in 2 episodes
• In International Conference on Machine Learning — 2× in this paper
• Provenance — 2× in this paper · explained in 4 episodes
• Attention Mechanism — 1× in this paper · explained in 3 episodes
• CVF International Conference on Computer Vision — 1× in this paper
• Fine-Tuning — 1× in this paper · explained in 5 episodes
• International Conference on Learning Representations — 1× in this paper
• KV cache — 1× in this paper · explained in 3 episodes
• Quantization-Aware Training — 1× in this paper · explained in 3 episodes
• Verification — 1× in this paper · explained in 2 episodes

Sources

• INT-FlashAttention: Enabling Flash Attention for INT8 Quantization
• FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
• ITA: An Energy-Efficient Attention and Softmax Accelerator for Quantized Transformers
• I-BERT: Integer-only BERT Quantization
• I-ViT: Integer-only Quantization for Efficient Vision Transformer Inference
• LRP-QViT: Mixed-Precision Vision Transformer Quantization via Layer-wise Relevance Propagation
• FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
• P$^2$-ViT: Power-of-Two Post-Training Quantization and Acceleration for Fully Quantized Vision Transformer
• EXAQ: Exponent Aware Quantization For LLMs Acceleration
• VFA: Relieving Vector Operations in Flash Attention with Global Maximum Pre-computation
• Efficient Attention Mechanisms for Large Language Models: A Survey
• A Survey on Transformer Compression
• MXAttention: Data-Free Optimal Scaling and Pre-Normalization Quantization for MXFP4 Attention
• PTQ4ViT: Post-training quantization for vision transformers with twin uniform quantization
• FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling
• SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization
• SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration
• SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training
• IntAttention: A Fully Integer Attention Pipeline for Efficient Edge Inference

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks