jt-zhang

/

SageAttention3

like 32

Current status – build still fails (tested with both C++17 and C++20)

Environment
• Windows 11
• Python 3.13.5
• Visual Studio 2022 Build Tools 17.10 (MSVC 19.40.33813)
• CUDA 12.8
• PyTorch 2.8.0 + cu128
• RTX-5090 (Blackwell sm_120)

What I tried and what happened

Default pip install -e .
→ Build stops in CUTLASS with
cutlass::platform::is_unsigned_v not found.
(Cause: the project is compiled with /std:c++17, which is too low.)

Explicitly forced /std:c++17 again (env vars)
→ Same error. Confirms C++17 cannot compile current code.

Switched to /std:c++20 and set DISTUTILS_USE_SDK=1
→ Build script still appends /std:c++17; MSVC prints
D9025: '/std:c++17' overrides '/std:c++20' and we are back to step 1.

Edited setup.py – replaced every std=c++17 with C++20
→ D9025 is gone. CUTLASS now compiles, but build fails in
sageattn/blackwell/kernel_traits.h with

error C2061: syntax error: identifier 'LayoutSF'
error C2061: syntax error: identifier 'SfAtom'
warning C4346: dependent name is not a type

Open problems

1. The build script always forces C++17 flags; users cannot raise the standard.
2. kernel_traits.h is missing typename for two dependent types.