This is the release note of v13.5.0. See here for the complete list of solved issues and merged PRs.

💬 Join the Matrix chat to talk with developers and users and ask quick questions!

🙌 Help us sustain the project by sponsoring CuPy!

✨ Highlights

• CuPy now supports NVIDIA CUDA 12.9 and AMD ROCm 6.4 platforms, and NumPy 2.3.
• Unified Memory Programming support for HMM/ATS-enabled systems (such as NVIDIA Grace Hopper Superchip) has been added. Refer to the documentation for the usage.
• Binary packages on PyPI (wheels) can now load NCCL packages installed via Pip (e.g., nvidia-nccl-cu12). In addition, Arm (aarch64) wheels are now built with NCCL support enabled.

Request for Comments

We are going to finalize the following RFC issues.

• Drop support for cuDNN in CuPy v14 (#8215)
• Update set of supported ROCm versions in CuPy v13/v14 (#8607)
• Remove cupyx.tools.install_library in CuPy v14 (#9204)

📝 Changes

New Features

• Support system allocated memory (#9033)

Enhancements

• Fix rocThrust build for ROCm 6.3 (#9023)
• Allow discovering cuTENSOR using major version (#9037)
• Support FIPS enabled machines with MD5 hashing (#9055)
• Update cutensornet accelerator based on cuquantum-python 25.03 deprecation (#9058)
• Refactor hashing (#9059)
• Raise user warning in both {to,from}Dlpack & Update the Interoperability page (#9061)
• Allow build on ROCm 6.4 (#9100)
• Migrate to pyproject.toml (#9135)
• Support NCCL for aarch64 (#9141)
• Support loading NCCL from Pip packages (#9208)
• Support CUDA 12.9 and NCCL 2.26 (#9211)
• Fix cupyx.scipy.stats.zscore for SciPy 1.15 (#9024)

Performance Improvements

• Implement lazy load for cuquantum (#9104)

Bug Fixes

• JIT: Support empty return (#9001)
• API: Revert toDlpack() default to the old unversioned one (#9007)
• BUG: Hot fix for numpy 2 support in some fusion paths (#9012)
• Fix compilation error of cupy.inf in fusion2 (#9043)
• Support Cython 3.1 (#9132)
• Fix cupyx.scipy.linalg.expm (#9144)

Code Fixes

• Fix get_typename to emit thrust::complex (#9054)

Documentation

• Add an AI policy to prohibit misuse of the issue tracker (#9095)
• Update ROCm docs (#9108)
• Docs: Update build-time requirement of Cython (#9145)
• Fix WARNING: Inline emphasis start-string without end-string (#9168)
• Improve API reference list (#9189)
• Bump supported NumPy version to v2.3 (#9203)

Installation

• Limit Cython version to 3.0 or 3.1 (#9146)
• Bump NumPy version restriction (#9166)
• Make rebuild faster for development (#9196)
• Bump version to v13.5.0 (#9212)

Tests

• CI: Do not run full CI on CUDA 12.0/12.1/12.2 + Windows (#9000)
• CI: Pin setuptools version on Windows (#9039)
• Revert "CI: Pin setuptools version on Windows" (#9056)
• Mark xfails in some spline tests for SciPy 1.15 (#9060)
• Support SciPy 1.15 (#9063)
• Skip some dtype checks with NumPy 2.x (#9064)
• Skip tests for different behavior of integer overflow from NumPy 2 (#9072)
• Skip some cupyx.scipy.special tests for SciPy 1.15 (#9073)
• Skip some tests for numerical error from NumPy 2 (#9075)
• Do gc.collect() in MemoryHook test code to avoid free hook to happen (#9093)
• np.sum has numerical change in NumPy 2.3 (#9169)
• Fix cp.empty(None) to raise TypeError (#9174)
• CI: NumPy 2.3 (#9194)
• Add NumPy 2.3 + windows CI (#9197)
• Update pre-commit settings (#9202)
• Add cupy.win.cuda129 CI (#9214)
• Fix test trigger phrase for cupy.win.cuda129 CI (#9217)

Others

• Allow specifying no libraries when generating wheel metadata (#9080)
• Upgrade pre-commit hooks (#9156)

👥 Contributors

The CuPy Team would like to thank all those who contributed to this release!

@asi1024 @Azusachan @EarlMilktea @ev-br @jakirkham @kmaehashi @leofang @MattTheCuber @rongou @seberg @yangcal