fullseye

Building native gsplat on Windows (RTX 5090 / torch cu128) — a proven procedure

日本語 · English

On 2026-08-19, native gsplat was built successfully on an RTX 5090 (1532 rasterizations/s). This note records the key points and the fixes for the sticking points.

Environment

• GPU: RTX 5090 (Blackwell sm_120) / driver 610.74
• venv: <local working path>\dev\projects\imgevolve\.venv-gsplat (torch 2.11.0+cu128, gsplat 1.5.3)
• The shared py -3.11 was left unchanged

Required toolchain

1. MSVC C++ (cl.exe): “Desktop development with C++ (VCTools)” from VS BuildTools 2022. Installation requires administrator elevation (UAC). ...\BuildTools\VC\Tools\MSVC\14.44.35207\bin\Hostx64\x64\cl.exe
2. nvcc 12.8 (matching the version of torch cu128): System CUDA 13.3 is header-incompatible with torch 2.11. Obtain it via conda, which requires no admin:

   micromamba create -p <prefix>\cuda128 -c conda-forge "cuda-nvcc=12.8.*" "cuda-cudart-dev=12.8.*" "cuda-cccl=12.8.*" "cuda-nvrtc-dev=12.8.*"
   # CUDA_HOME = <prefix>\cuda128\Library, nvcc = Library\bin\nvcc.exe
   # torch looks at lib/x64, so: cp Library/lib/*.lib Library/lib/x64/
   

Build-time environment (bat)

call "...\BuildTools\VC\Auxiliary\Build\vcvars64.bat"
set CUDA_PATH=<prefix>\cuda128\Library
set CUDA_HOME=%CUDA_PATH%
set PATH=%CUDA_PATH%\bin;%CUDA_PATH%\nvvm\bin;%PATH%
.venv-gsplat\Scripts\python.exe -c "import gsplat, torch; gsplat.rasterization(...)"  # JIT build on first run (~44s)

Applied patches (★ they vanish on reinstall. They must be reapplied)

1. small macro clash in the torch header (torch 2.11 Windows bug): In .venv-gsplat\Lib\site-packages\torch\include\c10\cuda\CUDACachingAllocator.h, insert #ifdef small / #undef small / #endif immediately before struct StreamSegmentSize. (Because Windows rpcndr.h’s #define small char turns bool small into bool char.)
2. Remove gsplat’s -Wno-attributes under MSVC: .venv-gsplat\Lib\site-packages\gsplat\cuda\_backend.py extra_cflags = [opt_level] if os.name=="nt" else [opt_level, "-Wno-attributes"] (cl.exe rejects GCC-style flags with D8021.)

Result

gsplat.rasterization works, 50 iterations in 0.033s = 1532/s. It can be swapped in as a backend that is orders of magnitude faster than pure torch (non-tiled, gsplat_torch.py).

This site is open source. Improve this page.