CUDA Python has reached a 1.0 milestone alongside CUDA 13.3, consolidating cuda.core, cuda.compute, cuda.bindings, cuda-pathfinder, and nvmath-python into a single officially supported, NVIDIA-maintained way to reach the CUDA platform from Python. The release commits these components to semantic versioning: breaking changes only in major releases, deprecation before removal, and predictable upgrade paths. cuda.core turns CUDA's runtime concepts (devices, streams, buffers) into standard Python objects, giving libraries like Numba, cuda.compute, CuPy, PyTorch, NCCL4Py, and NVSHMEM4P a shared foundation so they can interoperate on the same GPU memory without copying. New cuda.core capabilities include green contexts for SM partitioning, process checkpointing, and inter-process memory sharing. Installation is a single pip command, and getting started guidance is organized by tier: RAPIDS for data science, cuda.compute for optimized algorithms, Numba/Numba CUDA MLIR for custom kernels, and cuda.core/cuda.bindings for low-level driver access.
Table of contents
CUDA 1.0: Semantic versioningOne foundation instead of manyThe mental model: Three tiers on one foundationThree ways inThe ecosystem is already convergingGetting startedQuestions this post answers
What is included in the CUDA Python 1.0 release with CUDA 13.3?
CUDA Python 1.0 bundles cuda.core 1.0.0 for Pythonic CUDA runtime access, cuda.compute 1.0.0 for CCCL parallel algorithms, cuda.bindings 13.3.0 for low-level 1:1 CUDA C API bindings, cuda-pathfinder for locating installed CUDA components, and nvmath-python 1.0 for NVIDIA's math libraries. Each component is versioned independently rather than sharing the 1.0 number, and the milestone commits them to semantic versioning with deprecation before removal. Developers standardizing GPU stacks can track how each CUDA Python component evolves through daily.dev.
How do I install the CUDA Python 1.0 stack with pip?
Run pip install cuda-python cuda-cccl numba-cuda-mlir[cu13] to get cuda.core, cuda.compute, cuda.bindings, cuda-pathfinder, and the MLIR-based Numba backend. Install nvmath-python separately with pip install nvmath-python[cu13]. Only an up-to-date NVIDIA driver is required; a separate CUDA Toolkit installation is generally not needed. Anyone setting up a GPU-accelerated Python environment can follow updates like this via daily.dev.
What are green contexts in cuda.core and why do they matter?
Green contexts let a GPU's streaming multiprocessors be partitioned into disjoint groups so latency-sensitive kernels are shielded from long-running throughput kernels running in the same process. Previously, each library that wanted this had to bind and expose the capability independently; with cuda.core 1.0 it is implemented once and any library built on cuda.core can use it, alongside process checkpointing and inter-process GPU memory sharing. Engineers tuning GPU workload isolation can follow deep dives on features like this through daily.dev.