v3.2.0-windows.post13

TinyCC is bundled in the wheels, so you don't need to install MSVC to use Triton. Packages that directly call triton.jit, such as SageAttention, will just work.

You still need to install a C++ compiler if you use torch.compile targeting CPU. This may happen when you use nodes like 'CompileModel' in ComfyUI. Triton does not affect how PyTorch configures the C++ compiler in this case.

tcc

TinyCC is bundled in the wheels, so you don't need to install MSVC to use Triton.

TinyCC 0.9.27 is downloaded from https://download.savannah.gnu.org/releases/tinycc/tcc-0.9.27-win64-bin.zip

TinyCC 0.9.28rc is built from https://github.com/TinyCC/tinycc . It implements stdalign.h, which is needed since Triton 3.5, see triton-lang#7987 .

The def files used by Triton are generated by

tcc -impdef C:\Windows\System32\nvcuda.dll -o lib\cuda.def
tcc -impdef path\to\python3.dll -o lib\python3.def
tcc -impdef path\to\python39.dll -o lib\python39.def
...

For reference, the versions of these DLLs are

nvcuda.dll 32.0.15.7270
python3.dll 3.9.13150.1013
python39.dll 3.9.13150.1013
python310.dll 3.10.11150.1013
python311.dll 3.11.9150.1013
python312.dll 3.12.9150.1013
python313.dll 3.13.2150.1013
python314.dll 3.14.123.1013
python3t.dll 3.13.2150.1013
python313t.dll 3.13.2150.1013
python314t.dll 3.14.123.1013

The pip package tinycc was not used because these def files also need to be bundled.

v3.2.0-windows.post12

Let the environment variables TRITON_LIBCUDA_PATH and CUDA_PATH take higher precedence than the bundled CUDA

v3.2.0-windows.post11

• Since the release post11, the wheels are published to https://pypi.org/project/triton-windows/ , and no longer to GitHub. You can simply install the wheel using pip install -U triton-windows
• A minimal toolchain of CUDA is bundled in the wheels, so you don't need to manually install it. (You still need to manually install MSVC, Windows SDK, and vcredist)
• The wheels are linked against the LLVM from oaitriton.blob.core.windows.net, built by https://github.com/triton-lang/triton/blob/main/.github/workflows/llvm-build.yml , to better align with the official Triton
• The JIT-compiled C binaries (cuda_utils.pyd, __triton_launcher.pyd) are linked against the Python stable ABI, so there should be less error like DLL load failed while importing cuda_utils when switching the Python version

v3.2.0-windows.post10

For conda, support pytorch-gpu installed in conda-forge channel and cuda-toolkit installed in nvidia channel. Starting from PyTorch 2.6, PyTorch is no longer released in pytorch channel

v3.2.0-windows.post9

Following the official Triton, I release wheels for Python 3.9 to 3.13 .

v3.1.0-windows.post9

• Fix PTX ISA version for CUDA 12.8
• Fix int64 overflow in make_launcher

v3.1.0-windows.post8

Support CUDA from pip

v3.1.0-windows.post7

Fix build.py

v3.1.0-windows.post6

• Avoid creating files like main.obj in the working dir
• Refactor windows_utils.py to more thoroughly find MSVC and CUDA

Edit: Do not use this release, as there was a silly mistake