🧰 FlashAttention

The items that name it 9

• Folding Attention Biases into Keys Makes Compacted KV Caches Run on Standard Kernels
  paperas frameworks and tools2026-10-27

  Standard attention kernel that has no input for Attention Matching's per-key bias.

• Velox: Spatial Memory Makes World Models Efficient
  paperas frameworks and tools2026-10-15

  Tiled attention execution style that Velox's fixed token-block granularity is designed to match

• Visual Tokens on Demand: Visual Budgeting for Scalable and Efficient Multi-Image Reasoning in Multimodal LLMs
  paperas frameworks and tools2026-10-13

  Memory-efficient attention implementation that makes attention matrices impractical to access for token-importance scoring

• Academia is for Ambition — Alex Zhang, MIT
  writeupas classification.key entities2026-10-02
• The Roadmap to Mastering LLM Inference Optimization
  articleas techniques2026-09-21
• whitecircle/halo
  repoas frameworks and tools2026-08-26It is written here as Flash Attention.

  Attention kernel (FA2, FA3, FA4) by Tri Dao and contributors, used for attention and kernels in Halo, vendored as a git submodule.

• patchy631/time-to-first-token
  repoas classification.key entities2026-08-02
• deepbeepmeep/Wan2GP
  repoas frameworks and tools2026-07-30It is written here as Flash Attention.

  One of the acceleration kernels automatically installed by WanGP's install scripts.

• vLLM v0.30.0
  repoas frameworks and tools

  Attention kernel library, pin synced with upstream; used in DFlash drafters