Project Status

CuFlash-Attn is maintained as a stable v0.3.0 baseline and an archive-ready reference implementation. The current work focuses on governance cleanup, documentation quality, workflow simplification, and bug fixes rather than feature expansion.

What this project covers

• From-scratch CUDA C++ implementation of FlashAttention
• Forward and backward passes for float and half
• Supported head_dim values: 32, 64, 128
• Public C++ API plus C ABI for Python ctypes integration
• OpenSpec-tracked design and verification rules

Maintenance posture

This repository is intentionally optimized for:

• clarity over breadth: no speculative feature growth
• stable integration surface: examples, docs, and ABI stay aligned
• lightweight engineering: preset-based builds, focused CI, bilingual docs
• handoff readiness: contributors and follow-up models can continue from explicit specs and control docs

Development workflow

The preferred workflow is:

1. Read the relevant files in openspec/specs/
2. Capture changes through an OpenSpec change when scope or behavior shifts
3. Build with CMake presets only
4. Run verification appropriate to the environment
5. Use review before landing non-trivial changes

Validation boundaries

• Local CUDA builds require a working toolkit and nvcc
• GPU tests are skipped automatically on systems without a CUDA device
• Documentation and workflow cleanup can be validated without a GPU

Canonical references

• Quick Start
• Building from Source
• API Reference
• Troubleshooting
• CHANGELOG.md
• OpenSpec Specifications