Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

49 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

FlashAttention Notes

Source-backed learning notes for FlashAttention 1 through 4: the attention math, tile and worker ownership, memory movement, GPU execution primitives, performance rationale, and determinism boundaries.

The material is organized as a progressive lowering path:

attention math
  -> online softmax and IO-aware tiling
  -> CTA / warp ownership
  -> registers / SMEM / TMEM / HBM
  -> MMA, copy, barrier, and scheduling primitives
  -> performance and determinism consequences

Published Site

Configure GitHub Pages to publish from:

branch: main
folder: /docs

The site will then be available at:

https://zyeric.github.io/fa-notes/

Start Here

Reading Surfaces

The Markdown files under docs/notes/ are the source of truth. They pin papers, source revisions, implementation landmarks, inference boundaries, and future GPU checks.

The standalone HTML decks under docs/slides/ are visual reading surfaces:

Generation Forward Backward Markdown source
FA1 / Ampere slides slides foundations
FA2 / Ampere slides slides forward, backward
FA3 / Hopper combined slides same deck deep dive
FA4 / Blackwell combined slides same deck deep dive
Large-d inference schedule comparison out of scope source-backed map

Rubin is tracked separately as an architecture projection, not labeled FA5.

docs/notes.html is a generated long-form HTML view for readers who prefer a browser over GitHub Markdown.

Repository Layout

docs/
  index.html                 # GitHub Pages landing page
  notes.html                 # generated long-form reading surface
  render_notes.py            # dependency-free renderer
  notes/                     # Markdown source of truth
  slides/                    # standalone visual decks
STATUS.md                    # completion and validation boundary
PROVENANCE.md                # extraction and source-history record

Scope

The first pass covers ordinary training attention and the historical evolution from FA1 through FA4, plus a clearly labeled Rubin projection. A narrow large-head-dimension inference-forward extension compares public DeepSeek, FlashMLA, vLLM, SGLang, and FA4 evidence without turning it into a universal backend verdict. Full paged/decode qualification, inference-engine scheduling, large-d training backward, batch invariance, and linear attention remain separate work.

Many physical and performance conclusions are source-backed but still CPU-only. The documents label where SASS inspection, profiling, or repeated GPU execution remains necessary.

Regenerating The Long-Form Page

python3 docs/render_notes.py

The renderer uses only the Python standard library. The standalone slides are maintained directly and are not generated from Markdown.

Provenance

This repository was split from the FlashAttention learning context originally maintained in axis-training-dev-tools. Relevant commit history was retained; see PROVENANCE.md.

License

License is not selected yet. Choose an explicit content/code license before promoting broad reuse.

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors