How to Install Triton

Triton is OpenAI’s open-source compiler for writing GPU kernels in Python. Most users encounter it as a transitive dependency of PyTorch, but it can also be installed standalone for custom kernel development. Triton only ships wheels for Linux (x86_64 and aarch64). There are no macOS or Windows builds.

Requirements

• Linux (x86_64 or aarch64). Triton has no wheels for macOS or Windows.
• Python 3.10 or later (up to 3.14).
• glibc 2.27 or later (Ubuntu 18.04+, Debian 10+, RHEL 8+).
• An NVIDIA or AMD GPU with appropriate drivers for running compiled kernels. Triton itself installs without a GPU, but kernels cannot execute without one.

Install as a PyTorch dependency (automatic)

PyTorch declares Triton as a dependency on Linux x86_64. Installing PyTorch pulls in the matching Triton version automatically:

uv add torch

Or with pip:

pip install torch

Each PyTorch release pins a specific Triton version:

No extra configuration is needed. The version constraint in PyTorch’s metadata ensures the correct Triton version gets installed.

Install standalone

Triton is published to PyPI under the package name triton:

uv add triton

Or with pip:

pip install triton

The standalone triton package on PyPI is the same package that PyTorch depends on. There is no separate “pytorch-triton” package to worry about (the pytorch-triton package on PyPI is an empty placeholder).

Install a specific version

Pin the version to match a PyTorch release or to get a specific feature:

uv add "triton==3.5.1"

When Triton is installed alongside PyTorch, the versions must be compatible. Installing a Triton version that conflicts with PyTorch’s pinned requirement will cause a resolution error. If that happens, either remove the explicit Triton pin or upgrade PyTorch to match.

Install with conda-forge or pixi

Triton is available on conda-forge for Linux (x86_64 and aarch64):

conda install -c conda-forge triton

Or with pixi:

pixi add triton

conda-forge tracks Triton releases closely and packages versions 3.1.0 through 3.6.0. This can be a good option when managing CUDA dependencies through conda, since conda resolves the CUDA toolkit as part of the same dependency graph (see choosing between uv and conda for scientific Python).

Verify the installation

Confirm Triton is installed and can compile a kernel:

python -c "import triton; print(triton.__version__)"

To verify that GPU compilation works, run a minimal kernel on a machine with a GPU:

import torch
import triton
import triton.language as tl

@triton.jit
def add_kernel(x_ptr, y_ptr, out_ptr, n, BLOCK: tl.constexpr):
    idx = tl.program_id(0) * BLOCK + tl.arange(0, BLOCK)
    mask = idx < n
    x = tl.load(x_ptr + idx, mask=mask)
    y = tl.load(y_ptr + idx, mask=mask)
    tl.store(out_ptr + idx, x + y, mask=mask)

x = torch.randn(1024, device="cuda")
y = torch.randn(1024, device="cuda")
out = torch.empty_like(x)
add_kernel[(1,)](x, y, out, 1024, BLOCK=1024)
assert torch.allclose(out, x + y)
print("Triton kernel executed successfully")

Troubleshooting

triton version conflicts with PyTorch. PyTorch pins an exact Triton version (e.g., triton==3.5.1). If a project also pins a different version, the resolver will fail. Remove the explicit Triton pin, or upgrade PyTorch to a release that matches the desired Triton version.

No wheels found (macOS or Windows). Triton only publishes Linux wheels. On macOS and Windows, pip install triton will fail with a “no matching distribution” error. Triton cannot be used on these platforms. For development workflows that need cross-platform support, guard the dependency with an environment marker:

[project]
dependencies = [
    "triton>=3.5.0; sys_platform == 'linux'",
]

GLIBC version too old. Triton wheels require glibc 2.27+. On older Linux distributions, the install will fail with an error about an incompatible platform tag. Upgrade the OS or use a container with a newer glibc.

Kernel compilation fails at runtime. Triton compiles kernels to PTX/AMDGPU IR at runtime, which requires a working GPU driver. If the CUDA or ROCm driver is missing or misconfigured, compilation will fail even though the package installed correctly. Verify driver installation with nvidia-smi (NVIDIA) or rocm-smi (AMD).

Related

• How to Install PyTorch with uv
• Why Installing GPU Python Packages Is So Complicated
• uv vs. pixi vs. conda for Scientific Python