AINewsnow

Low Precision Flash Attention 4: End-to-End Block-Scaled Attention for Blackwell

This story is from 2026-09-16. It is preserved in the archive; the latest stories are on the live feed.

TL;DR We extend FlashAttention-4 [1] with MXFP8 forward and backward, reaching 2.85 PF/s forward and 2 PF/s backward on LLM shapes. On our internal shapes, FA4 MX8 reaches 2.54 PF/s...

Read the full story at PyTorch Blog ↗

Timeline · 1 report

  1. 2026-09-16 18:55 · PyTorch Blog
    Low Precision Flash Attention 4: End-to-End Block-Scaled Attention for Blackwell

More stories

  1. Google Joins OpenAI, Anthropic, Meta in Disclosing AI Hacks — Bloomberg AI
  2. Introducing Kimi K3 on Amazon Bedrock — AWS Machine Learning Blog
  3. Introducing Amazon SageMaker HyperPod Inference Gateway — AWS Machine Learning Blog
  4. Introducing Astra for Law — OpenAI News
  5. Newsom signs executive order to explore new AI rules, consider ‘kill switch’ — Politico Technology
  6. Anthropic, OpenAI, SpaceXAI, Google sued over call to ‘pace’ AI development — Politico Technology
  7. Sources: Anthropic considers releasing a new AI model to counter OpenAI's momentum since Astra's launch, ahead of an IPO and after Amodei's call for a slowdown (Reuters) — Techmeme
  8. Gemini Hacked Three Companies in First Known Breakout by Google’s AI — Wall Street Journal Technology

Get the daily brief of stories like this at 6:30 every morning →