Format-Aware Fusion for Fast FP4 Pretraining
cs.LG
Submitted: 2026-09-04
Updated: 2026-09-04
Code: https://github.com/MrHuff/mfu-fp4-training
Sources
- QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs
- TetraJet-v2: Accurate NVFP4 Training for Large Language Models with Oscillation Suppression and Outlier Control
- Accurate Neural Training with 4-bit Matrix Multiplications at Standard Formats
- Pretraining large language models with MXFP4 on Native FP4 Hardware
- Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling
- FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning
- FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
- Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention
- Dissecting Outlier Dynamics in LLM NVFP4 Pretraining
- The Llama 3 Herd of Models
- Is Finer Better? The Limits of Microscaling Formats in Large Language Models
- CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs
- Measuring Massive Multitask Language Understanding
- Liger Kernel: Efficient Triton Kernels for LLM Training
- Elucidating the Design Space of FP4 training
- UE5M3 FP4 Block Scaling for Stable Language Model Pretraining
- FP8 Formats for Deep Learning
- Pretraining Large Language Models with NVFP4
- Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation
- Stable FP4 Training via Transposition-Invariant Block Quantization
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: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks