tokenspeed-triton

What it is and what it does

Triton is a compiler and language designed to let you write high-performance deep-learning kernels at a higher level of abstraction than CUDA, while retaining the ability to optimize for specific hardware. Instead of writing low-level GPU code, you write Triton kernels in Python-like syntax, and the compiler handles the translation to efficient machine code for GPUs and CPUs. It sits between the productivity of high-level frameworks and the control of hand-written CUDA.

The package is intended for researchers and engineers building custom neural-network operations—operations that existing frameworks don't provide or that need domain-specific optimization. It includes a just-in-time compiler, an interpreter for CPU-based testing, and support for tiled computation patterns common in deep learning. The main runtime dependency is importlib-metadata; the package itself handles LLVM integration during build time.

Use it for

• Write custom CUDA-like kernels for novel neural-network layers without learning low-level GPU programming.
• Optimize matrix operations, attention mechanisms, or other primitives for specific hardware without rewriting in C++.
• Prototype and test GPU kernels on CPU using the Triton interpreter before deploying to hardware.
• Build domain-specific deep-learning libraries that need fine-grained control over memory and compute.
• Accelerate research by rapidly iterating on custom operations in a higher-level language than CUDA.

Worth the install?

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

Install

tokenspeed-triton on PyPI

Before you install

Medium install friction: wheels are available for Python 3.10–3.14 on x86_64 and aarch64 Linux, but building from source requires LLVM and build-time dependencies. The package is actively maintained with a recent release (24 days old) and a well-established repository.

Requires a compatible GPU (NVIDIA or AMD) and CUDA/ROCm runtime for execution; CPU-only testing is possible with TRITON_INTERPRET=1 environment variable.

License in practice

MIT license is permissive; you can use, modify, and distribute this package with minimal restrictions, provided you include the license notice.

Quickstart

pip install tokenspeed-triton

import triton
import triton.language as tl

@triton.jit
def kernel(x_ptr, y_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)
    y = x * 2
    tl.store(y_ptr + offsets, y, mask=mask)

Package facts

Evidence: tokenspeed_triton-3.8.10.post20260721-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl; tokenspeed_triton-3.8.10.post20260721-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl; tokenspeed_triton-3.8.10.post20260721-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl; tokenspeed_triton-3.8.10.post20260721-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl; tokenspeed_triton-3.8.10.post20260721-cp312-abi3-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl; tokenspeed_triton-3.8.10.post20260721-cp312-abi3-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

Tags

More Build Tools packages

See also triton · triton-windows · triton-ascend · numba · helion · liger-kernel · perf-analyzer · flash-attn · nvidia-nvvm · pyston

Further reading