CuPy

CuPy v14: NumPy v2 Semantics, bfloat16, CUDA Pip Wheels Support, and More! 🚀 We are thrilled to announce the release of CuPy v14, a major milestone in our journey to make GPU computing more accessible and interoperable for the Python community. Coming on the heels of our 10th anniversary in 2025—and surpassing 60 million total downloads—this release represents our first major update in two years. 🔢 Full NumPy v2 Alignment: Updated semantics and type promotion rules to match the latest NumPy 2 specification. ✨ Extended Data Types: Introduced initial support for bfloat16 and structured dtypes. 🐍 CUDA Pip Wheels Support: Deployment just got easier! You can now install CuPy along with the necessary CUDA Toolkit components directly via pip, without needing a system-wide CUDA installation. 📐 Enhanced API Coverage: Over 50 new APIs added, including key linear algebra routines (cupy.linalg.eig) and SciPy interpolation routines. We are also proud to welcome two long-time contributors, Leo Fang and Sebastian Berg, as our newest maintainers. Their expertise from the broader Python and CUDA ecosystems will be invaluable as we head into our second decade. 🔗 Read the full release blog here: https://lnkd.in/gEyTBCYy A huge thank you to our 10,000+ GitHub stargazers and every contributor who made this release possible. We can’t wait to see what you build with CuPy v14! ⚡

CuPy v14.0.0rc1 (Release Candidate 1) is officially available for testing! We encourage you to try it out and share your feedback on our issue tracker as we prepare for the final v14 release in January 2026. ✅ CUDA PyPI Package Compatible: Run CuPy without a system-wide CUDA Toolkit installation! This also enables better interoperability with the broader CUDA + Python ecosystem like PyTorch. ✅ AMD ROCm 7 Support: Quick setup with the new binary package, cupy-rocm-7-0. ✅ Enhanced NumPy/SciPy API Coverage: Includes new functions like cupy.linalg.eig and cupy.linalg.eigvals. Read the full draft release note for all the details. Your testing and feedback are invaluable! 🗒️ Release Notes: https://lnkd.in/gZ3utrVW #CuPy #GPU #Python #CUDA #ROCm #DataScience

Big congrats to the scikit-learn team on the 1.8 release! We're especially excited about the expanded Array API support. With this update, you can pass CuPy arrays directly to even more estimators, which means more of scikit-learn's advanced features can now run faster on the GPU. ⚡️🔥 It's great to see the Python ML ecosystem becoming more interoperable! 🙌 #scikitlearn #CuPy #MachineLearning #opensource #DataScience #Python #ML

We are happy to announce the official release of CuPy v13.6! The highlight of this release is the introduction of support for NVIDIA CUDA 13, the latest major version of the CUDA Toolkit. This update allows you to harness the newest features, performance optimizations, and hardware support from NVIDIA for your GPU-accelerated workloads. Pre-compiled binary packages are immediately available for Linux (x86_64 & arm64) and Windows (x86_64). To install the specific package for CUDA 13, use the following pip command: `pip install cupy-cuda13x`. For a complete list of all enhancements and bug fixes, please review the full release notes. Release Notes: https://lnkd.in/gHWFqDtj

A talk on CuPy will be presented at the SciPy 2025 conference! https://lnkd.in/gCnMJZez In this session, long-time contributor Leo Fang will share his personal story based on his experience contributing to open-source projects. He will cover CuPy's journey and how its JIT compiler has made GPU programming more accessible. Also, there are more sessions and tutorials highlighting CuPy at SciPy 2025. Explore the full schedule at https://lnkd.in/gS8HPPGc ! #CuPy #SciPy2025 #GPU #Python #ScientificComputing #OpenSource #NumPy #SciPy

We just released CuPy v13.5! CuPy now supports CUDA 12.9, ROCm 6.4, and NumPy 2.3. This release also brings Unified Memory Programming support for NVIDIA’s Grace Hopper systems. Check out the full release notes for details: https://lnkd.in/gvxaPKdV