Release v0.5.1

CuFlash-Attn v0.5.1

Changes | 变更

• fix(ci): align package smoke sanitizer runtime (98e6f8e)
• fix(ci): skip comparison without optional Python deps (9427c80)
• release: prepare v0.5.1 (4e84b79)
• fix: stabilize CUDA validation and documentation (e1329d4)
• docs: point LEARNING_PATH link to meta repo; status stable (743923e)
• docs: normalize org links to open-infra-ai; fix dead Pages deploy gate (77502c7)
• docs: use aicl-lab GitHub org in public links (b7ae7bd)
• docs: check off completed ROADMAP items from phase 2 (e0862b4)
• docs: link portfolio landing repo (6f246c3)
• perf(forward): skip fully-future KV blocks in causal path (e1735b3)
• fix(forward): flatten grid.y batch*heads for >65535 launches (d144765)
• docs: point links to cuda-foundations (4e773dc)
• docs(cuflash): archive v1.0 development plan under docs/development (7d4c534)
• docs(cuflash): sync v0.5.0 code facts and BF16 API sections (4927a2b)
• feat: E3 FlashDecoding (Split-KV) decode 前向 —— 两阶段分块 + 跨块归约 (9f65df7)
• docs: D2 补 IN/OUT 项目边界 —— FA 前后向/多精度/causal/FlashDecoding 为 IN；GEMM 基础/Triton/推理运行时为 OUT (d9ab221)
• docs: B3 benchmark 刷新 —— 新增 RTX 3060 Laptop 本机实测快照（head_dim 64/128） (52c4bfd)
• fix: correct benchmark timing and performance metrics (6860cbc)
• fix: use absolute GitHub URL for ROADMAP link in docs site (f588314)
• docs: add ROADMAP with kernel-depth optimization plan (11eff87)
• docs: mark benchmark data as v0.4.0 snapshot and fix maintainer wording (cd39503)
• docs: align claims with the verified learning scope (807ff0a)
• build(deps): bump actions/configure-pages from 5 to 6 (#10) (af5b1e2)
• build(deps): update torch requirement in /tests/integration (#9) (04f7323)
• build(deps): bump actions/upload-pages-artifact from 3 to 5 (#8) (70a0082)
• build(deps): bump softprops/action-gh-release from 2 to 3 (#6) (2937fe0)
• build(deps): bump actions/cache from 4 to 6 (#5) (66eb121)
• build(deps): bump actions/deploy-pages from 4 to 5 (#4) (d155740)
• build(deps): bump markdown-it-mathjax3 from 4.3.2 to 5.2.0 in /docs (#3) (96ec822)
• chore: 移除 dependabot 配置，停止自动依赖更新 PR (f421632)
• chore(docs): remove orphan architecture doc, fix stale doc path (d1f202e)
• docs(site): drop English docs, flatten URL structure to root (0a8c351)
• docs(repo): replace English README with Chinese-only version (c578f9d)

Install | 安装

The attached archive contains the shared library, public headers,
and the CMake package config. Verify its integrity first:

sha256sum -c cuflash-attn-v0.5.1-linux-x86_64.tar.gz.sha256

Or build from source:

cmake --preset release && cmake --build --preset release

CuFlash-Attn v0.5.0

补齐与 2026-08-05 源码里程碑对应的缺失标签；提交内容未改。该版本加入 WMMA 前向、FP32 logsumexp ABI、共享内存 tiling 回退与 stream-ordered workspace，详见 CHANGELOG。

CuFlash-Attn v0.4.0

补齐与 2026-07-28 源码里程碑对应的缺失标签；提交内容未改。该版本引入 BF16、统一 dtype 模板、FlashAttention-2 延迟归一化与公开 kernel primitives，详见 CHANGELOG。