# Hardware-Aware FP4 FlashAttention-4

arXiv: https://arxiv.org/abs/2609.04105 · HTML: https://mlsanity.com/paper/2609.04105/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 10 of its 13 references that are in the map.
- First posted on arXiv 3 Sep 2026.

## Builds on (most cited first, 10 of 10)

- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [Masked Autoencoders Are Scalable Vision Learners](https://mlsanity.com/paper/2111.06377/) (2021)
- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [Wan: Open and Advanced Large-Scale Video Generative Models](https://mlsanity.com/paper/2503.20314/) (2025)
- [FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning](https://mlsanity.com/paper/2307.08691/) (2023)
- [FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision](https://mlsanity.com/paper/2407.08608/) (2024)
- [SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training](https://mlsanity.com/paper/2505.11594/) (2025)
- [FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling](https://mlsanity.com/paper/2603.05451/) (2026)
- [Attn-QAT: 4-Bit Attention With Quantization-Aware Training](https://mlsanity.com/paper/2603.00040/) (2026)
- [UE5M3 FP4 Block Scaling for Stable Language Model Pretraining](https://mlsanity.com/paper/2609.02846/) (2026)

## Built on by (most cited first, 2 of 2)

- [Broken Symmetry in BF16 Attention: Why FlashAttention Gradients Blow Up Late in Training](https://mlsanity.com/paper/2609.34272/) (2026)
- [Backward-State Policy Is Part of the Learning Algorithm](https://mlsanity.com/paper/2609.39813/) (2026)

## Abstract (arXiv)

Blackwell's 4-bit floating-point (FP4) tensor cores do not automatically make attention faster because softmax conversion and on-chip dependencies dominate once its matrix products shrink. We address this with Direct-P for noncausal inference and a causal path that passes the forward quantization directly into backward. Direct-P maps scores directly to FP4 probabilities and reaches up to 2.13× the bfloat16 (BF16) forward throughput on an NVIDIA GB200. The causal path reconstructs probabilities from saved quantized queries and keys and uses 8-bit floating-point (FP8) gradient operands, accelerating a complete single-GPU 8-billion-parameter update by up to 1.14×. Matched distributed training retains FP8 probabilities and values; every tested MXFP4 probability/value training trajectory diverges.