한국어EN
Inference Infrastructure & ServingZadouri, Shah, Thakkar, Dao et al.Tri Dao

FlashAttention-4

FlashAttention-4 — co-design for asymmetric hardware scaling

Why it matters

The latest generation of the kernel lineage, formally published at MLSys 2026

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2603.05451

Added in Week 35, 2026 · See that week

Same week

Mishra et al. publish SoK: Agentic Retrieval-Augmented GenerationPaperReasoning Models Struggle to Control their Chains of Thought publishedPaperUN scientific panel of 40 holds first meeting, elects Bengio and Ressa co-chairsGovernmentUS Supreme Court declines to hear Thaler v. PerlmutterPress

Nearby on the same axis

Hugging Face archives the text-generation-inference repositoryOfficialvLLM introduces Model Runner V2OfficialGoogle Cloud makes TPU7x generally available (GA)OfficialGoogle Cloud previews its 8th-generation TPUs at Cloud NextPress