triton-windows

What it is and what it does

Triton-windows is a Windows-native fork of the Triton compiler, which allows you to write high-performance GPU kernels in Python and compile them just-in-time for NVIDIA GPUs. It bridges the gap between Python's ease of use and the performance demands of deep learning by letting you write custom operations that standard libraries don't cover, then compiling them directly to GPU code at runtime.

The package is designed specifically for Windows environments where the upstream Triton project did not have native support. It works with torch.compile and Triton's @triton.jit decorator, and passes all unit tests. Installation requires careful environment setup—you must ensure your Python version, PyTorch version, and GPU driver are aligned according to the documented compatibility matrix, and the package provides wheels only for Python 3.10–3.14 on Windows x86-64.

Use it for

• Accelerate custom deep learning operations on Windows by compiling hand-written GPU kernels instead of relying on standard libraries.
• Optimize matrix multiplication, attention mechanisms, or other compute-intensive layers by writing Triton kernels tailored to your model's needs.
• Use torch.compile with Triton on Windows to automatically fuse and compile operations into efficient GPU code.
• Prototype and deploy GPU-accelerated algorithms on Windows without dropping to lower-level GPU programming or managing low-level GPU memory manually.

Worth the install?

AI-flagged interpretation of the facts on this page. Verify before relying on it.

Install

triton-windows on PyPI

Before you install

Medium install friction: platform-specific wheels for Windows (cp310–cp314), requires matching Python environment and GPU driver setup. Repository is archived; development has moved to https://github.com/triton-lang/triton-windows, but pip installation continues to work.

Requires Windows 10 or 11, NVIDIA GPU with supported compute capability, matching PyTorch version (e.g., PyTorch 2.7 requires Triton 3.3), and correct Python environment setup (embedded, system-wide, conda, or venv).

License in practice

Permissive license (MIT) means you can use this package freely in commercial and private projects without restriction.

Quickstart

pip install triton-windows
import triton
@triton.jit
def kernel(x_ptr, output_ptr, n_elements, BLOCK_SIZE: tl.constexpr):
    pid = tl.program_id(axis=0)
    block_start = pid * BLOCK_SIZE
    offsets = block_start + tl.arange(0, BLOCK_SIZE)
    mask = offsets < n_elements
    x = tl.load(x_ptr + offsets, mask=mask)
    tl.store(output_ptr + offsets, x, mask=mask)

Package facts

Evidence: triton_windows-3.7.1.post27-cp310-cp310-win_amd64.whl; triton_windows-3.7.1.post27-cp311-cp311-win_amd64.whl; triton_windows-3.7.1.post27-cp312-cp312-win_amd64.whl; triton_windows-3.7.1.post27-cp313-cp313-win_amd64.whl; triton_windows-3.7.1.post27-cp314-cp314-win_amd64.whl

Tags

More Build Tools packages

See also tokenspeed-triton · triton · triton-ascend · nvidia-cuda-nvrtc-cu12 · nvidia-cuda-nvcc · helion · nvidia-cuda-nvcc-cu12 · nvidia-cuda-nvrtc · tritonclient · perf-analyzer

Further reading