cuda-tile
CUDA Tile Compiler
Decision gist · record as of 2026-08-14
Yes, if you have a supported NVIDIA GPU (Ampere, Ada, or Blackwell) and need to write custom GPU kernels in Python. The package is actively maintained, has no known vulnerabilities, and offers a genuine productivity gain over hand-written CUDA C++. Install friction is moderate due to driver and toolkit requirements, but prebuilt wheels and optional tileiras bundling ease setup. Not suitable if you lack a supported GPU or cannot meet the driver/toolkit prerequisites.AI-flagged interpretation of the facts on this page — verify before relying
Before you install
- Requires NVIDIA Driver r580 or later, CUDA Toolkit 13.1+, and a supported GPU (Blackwell, Ampere, or Ada).
- C++17 compiler and CMake 3.18+ needed to build from source.
- Medium install friction due to compiled C++ extension and NVIDIA driver requirements.
License · maintenance · safety
Apache-2.0 (permissive) — Licensed under Apache-2.0 (permissive), allowing commercial and private use with minimal restrictions; suitable for most projects.
last release 2026-07-08 (37 days) · last repo commit 2026-08-14 · 2,126 stars
0 known vulnerabilities (OSV.dev, 2026-08-14) · 3,783,144 downloads/mo, #2,498 on PyPI
Alternatives
Verify before relying
pip install cuda-tile[tileiras]
import cuda.tile as ct
import typing_extensions
@ct.kernel
def vector_add_kernel(a, b, result):
block_id = ct.bid(0)
a_tile = ct.load(a, index=(block_id,), shape=(16,))
b_tile = ct.load(b, index=(block_id,), shape=(16,))
result_tile = a_tile + b_tile
ct.store(result, index=(block_id,), tile=result_tile)- Whether Hopper GPU support is available in versions after 1.5.0 (docs note it 'will be supported in coming versions')
- Performance characteristics and overhead compared to hand-written CUDA C++
- Compatibility with GPU array libraries and ML frameworks for practical workflows
What it is and what it does
cuTile Python is a compiler that lets you write GPU kernels in Python using a tile-based programming model, targeting NVIDIA GPUs. You write kernels decorated with @ct.kernel, define tile loads and stores, and launch them on the GPU via ct.launch(). It abstracts away low-level CUDA details while giving you fine-grained control over how data is tiled and computed in parallel.
The package requires a modern NVIDIA driver (r580+) and CUDA Toolkit 13.1+ to run. It compiles kernels using Tile IR and the tileiras compiler, which currently supports Blackwell and Ampere/Ada GPUs. You can install it with optional tileiras support built into your Python environment, or use a system-wide CUDA Toolkit installation. It has minimal runtime dependencies (only typing-extensions) and provides prebuilt wheels for Python 3.10–3.14 on Linux and Windows.
Use it for
- Write custom GPU kernels for dense linear algebra operations without hand-coding CUDA C++
- Develop tile-based GPU computations for scientific computing or machine learning on supported GPUs
- Prototype and optimize GPU algorithms in Python with direct control over tile shapes and memory patterns
- Integrate GPU kernels into Python workflows for data interchange with GPU array libraries
Worth the install?
AI-flagged interpretation of the facts on this page. Verify before relying on it.
Yes, if you have a supported NVIDIA GPU (Ampere, Ada, or Blackwell) and need to write custom GPU kernels in Python.
The package is actively maintained, has no known vulnerabilities, and offers a genuine productivity gain over hand-written CUDA C++. Install friction is moderate due to driver and toolkit requirements, but prebuilt wheels and optional tileiras bundling ease setup. Not suitable if you lack a supported GPU or cannot meet the driver/toolkit prerequisites.
Install
cuda-tile on PyPI
Before you install
Medium install friction due to compiled C++ extension and NVIDIA driver requirements. Active maintenance with recent releases; requires NVIDIA Driver r580 or later and CUDA Toolkit 13.1+. Prebuilt wheels available for Python 3.10–3.14 on Linux (aarch64, x86_64) and Windows.
Requires NVIDIA Driver r580 or later, CUDA Toolkit 13.1+, and a supported GPU (Blackwell, Ampere, or Ada). C++17 compiler and CMake 3.18+ needed to build from source.
License in practice
Licensed under Apache-2.0 (permissive), allowing commercial and private use with minimal restrictions; suitable for most projects.
Quickstart
pip install cuda-tile[tileiras]
import cuda.tile as ct
import typing_extensions
@ct.kernel
def vector_add_kernel(a, b, result):
block_id = ct.bid(0)
a_tile = ct.load(a, index=(block_id,), shape=(16,))
b_tile = ct.load(b, index=(block_id,), shape=(16,))
result_tile = a_tile + b_tile
ct.store(result, index=(block_id,), tile=result_tile)Verify before relying
- Whether Hopper GPU support is available in versions after 1.5.0 (docs note it 'will be supported in coming versions')
- Performance characteristics and overhead compared to hand-written CUDA C++
- Compatibility with GPU array libraries and ML frameworks for practical workflows
Package facts
| License | Apache-2.0 permissive |
| Python support | Supports the current Python release <3.15,>=3.10 |
| Install friction | Medium. Platform-specific wheel |
| Runtime dependencies | 1 packagetyping-extensions |
| Maintenance | Actively maintained 37 days since the last release |
| Last repo commit | |
| First released | |
| Downloads | 3,783,144 / month, #2,498 on PyPI 30-day window, as of 2026-08-14 |
| Known vulnerabilities | None known OSV.dev, checked 2026-08-14 |
| Classifiers | Development Status :: 5 - Production/StableEnvironment :: GPU :: NVIDIA CUDA :: 13Intended Audience :: DevelopersProgramming Language :: Python :: 3.10Programming Language :: Python :: 3.11Programming Language :: Python :: 3.12Programming Language :: Python :: 3.13Programming Language :: Python :: 3.14Programming Language :: Python :: Implementation :: CPythonTopic :: Software Development :: Compilers |
Evidence: cuda_tile-1.5.0-cp310-cp310-manylinux2014_aarch64.whl; cuda_tile-1.5.0-cp310-cp310-manylinux2014_x86_64.whl; cuda_tile-1.5.0-cp310-cp310-win_amd64.whl; cuda_tile-1.5.0-cp311-cp311-manylinux2014_aarch64.whl; cuda_tile-1.5.0-cp311-cp311-manylinux2014_x86_64.whl; cuda_tile-1.5.0-cp311-cp311-win_amd64.whl; cuda_tile-1.5.0-cp312-cp312-manylinux2014_aarch64.whl; cuda_tile-1.5.0-cp312-cp312-manylinux2014_x86_64.whl; cuda_tile-1.5.0-cp312-cp312-win_amd64.whl; cuda_tile-1.5.0-cp313-cp313-manylinux2014_aarch64.whl; cuda_tile-1.5.0-cp313-cp313-manylinux2014_x86_64.whl; cuda_tile-1.5.0-cp313-cp313-win_amd64.whl; cuda_tile-1.5.0-cp314-cp314-manylinux2014_aarch64.whl; cuda_tile-1.5.0-cp314-cp314-manylinux2014_x86_64.whl; cuda_tile-1.5.0-cp314-cp314t-manylinux2014_aarch64.whl; cuda_tile-1.5.0-cp314-cp314t-manylinux2014_x86_64.whl; cuda_tile-1.5.0-cp314-cp314t-win_amd64.whl; cuda_tile-1.5.0-cp314-cp314-win_amd64.whl
Tags
Let your AI agent find packages like this
Example. Real query, live index.
You found this page by searching. An agent finds it by wishing: SkillFed indexes 14,416 PyPI packages by what they can do, searchable in plain language.
wish › “NVIDIA CUDA tile programming”
- cuda-tilecuTile Python is a compiler and programming language for writing GPU…
- nvidia-cutlass-dsl-libs-baseProvides Python interfaces for writing high-performance CUDA kernels…
- quadrantsQuadrants is a high-performance compiler that translates Python code…
Give your agent the search over MCP, or paste the wish link into any chat.
More Compilers packages
pyparsing provides a library for building text parsers directly in Python code using composable grammar classes, handling quoted strings, whitespace variation, and embedded comments without regex or lex/yacc.
Install it if you need to parse text or define grammars programmatically.
Cython is a source code translator that compiles Python code to C, enabling optional static typing and direct C/C++ interoperability to generate efficient native extensions.
Install it if you need to optimize Python performance, wrap C/C++ libraries, or build fast native extensions.
Annotates Python abstract syntax trees with the positions of tokens and source code text, enabling tools to map AST nodes back to their originating code for refactoring, analysis, and transformation.
Numba is a just-in-time compiler that accelerates numerically-focused Python code by compiling it to machine code via LLVM, with support for NumPy functions, loop parallelization, and GPU code generation.
Python bindings to the tree-sitter parsing library, enabling incremental parsing and syntax tree analysis for source code.
llvmlite provides a lightweight Python binding to LLVM's IR builder, optimizer, and JIT compiler APIs, enabling you to write JIT compilers and code generators without needing the full LLVM C++ API surface.
See also cuda-python · tilelang · nvidia-cutlass-dsl · nvidia-cutlass-dsl-libs-cu13 · nvidia-cutlass-dsl-libs-base · nvidia-cutlass-dsl-libs-cu12 · cuda-core · cufile-python · nvidia-cutlass-dsl-libs-core · cuda-bindings