Project Status

CuFlash-Attn is maintained as a stable v0.3.0 reference implementation for learning, auditing, and lightweight integration.

What stays in scope

• CUDA C++ FlashAttention from scratch
• Forward and backward passes for float and half
• Supported head_dim values: 32, 64, 128
• Public C++ API and C ABI for ctypes-style integration
• Bilingual technical documentation and GitHub Pages publishing

Maintenance posture

This repository now prefers:

1. clarity over process
2. deletion over framework sprawl
3. stable behavior over speculative features
4. one canonical source per topic

Canonical references

• Quick Start
• API Reference
• Architecture
• Repository README
• CHANGELOG.md

Validation boundaries

• Full CUDA builds require a working toolkit and nvcc
• GPU tests may be unavailable on documentation-only environments
• Docs and workflow cleanup should still keep the docs build and repository layout coherent