v26.06.00

This is a beta release of cuPyNumeric.

Pip wheels are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/ and https://pypi.org/project/nvidia-cupynumeric-cu12/, for Linux (x86-64 and ARM64, with CUDA 12/13 and single-node multi-GPU support) and macOS (for ARM64). Conda packages are available at https://anaconda.org/legate/cupynumeric, for Linux (x86-64 and ARM64, with CUDA 12/13 and multi-node support). Windows is currently supported through WSL.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/26.06/.

Support matrix changes

• Add CUDA 13 support to pip wheels
• Add support for Python 3.14 to all packages
• Add macOS pip wheel for the Legate profiler
• Remove networking support from pip wheels - due to limitations of the wheel distribution channel, the networking functionality of the wheels was not robust; use conda packages or build from source instead

Highlights

New features

• cupynumeric.linalg.cho_factor
• cupynumeric.linalg.cho_solve
• cupynumeric.linalg.inv
• cupynumeric.linalg.solve_triangular
• cupynumeric.mgrid
• cupynumeric.ndimage.convolve

Performance improvements

• Fix parallelization of cupynumeric.pad
• Improved Thrust- and NCCL-based implementations for advanced indexing operations
• Merge index array zipping with subsequent scatter/gather kernel in advanced indexing operations (single-GPU only)
• Add better heuristics for picking between cuSolver getrf/s and batched cuBLAS APIs
• Avoid some extraneous type conversions in reduction operations
• Reduce some Python overheads in ufunc operations

Examples

• Add a microbenchmark suite (ufunc, gemm, gemv, sort, reduction, indexing, FFT, etc.)
• Add examples showcasing different options for interoperating with PyTorch
• Add CFD example based on https://github.com/barbagroup/CFDPython

Tooling

• Add an Nsight Systems recipe that measures the degree of task-level parallelism
• Add a tutorial on profiling & debugging
• Add more anti-patterns to "cuPyNumeric Doctor" detector tool

UX improvements

• Remove fallback to NumPy for missing APIs and small arrays
• Add an option to disable bounds checking in various operations, which removes a source of blocking
• initial support for using cuPyNumeric ndarrays inside a Python Legate task (creation of task-local ndarrays is not supported yet)
• Initial implementation of the Python Array API

Known issues

• As of October 2025, Perlmutter jobs that request more than 32 GB of device memory (for example, --fbmem 64000) must include REALM_DEFAULT_ARGS='-gex:bindcuda 0'. Otherwise the OFI provider aborts with Unexpected error 12 (Cannot allocate memory) from fi_mr_regattr().
• We are aware of performance regressions with cupynumeric.einsum on Blackwell GPUs, starting to occur with cuBLAS 13.2. These are under investigation.
• With recent versions of UCX you might see warning messages like these:

  ib_md.c:296  UCX  ERROR ibv_reg_mr(address=(nil), length=134217728, access=0xf) failed: Bad address
  ucp_mm.c:81   UCX  ERROR failed to register address (nil) (cuda) length 134217728 on md[6]=mlx5_0: Input/output error (md supports: host|cuda)
  0.000000 {5}{ucp}: ucp_mem_map failed
  

  These can be ignored (they are not fatal, and don't appear to have an effect on performance). We are investigating how they can be addressed / removed.
• We are aware of possible hangs when calling APIs that use cuSolverMp (e.g. multi-GPU cupynumeric.linalg.solve). We are in contact with the cuSolverMp maintainers to address these. In the meantime, depending on the underlying cause, one or more of the following workarounds should resolve the hangs:
  • export NCCL_PXN_DISABLE=1
  • export CUDA_MODULE_LOADING=EAGER
  • run in rank-per-gpu mode

Known issues

This release is missing CUDA 13 wheels for Python 3.14 (conda packages for this combination are available). We are working to add this combination.

Full Changelog: v26.01.00...v26.06.00

v26.01.00

This is a beta release of cuPyNumeric.

Pip wheels are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/, for Linux (x86-64 and ARM64, with CUDA 12 and multi-node support) and macOS (for ARM64). Conda packages are available at https://anaconda.org/legate/cupynumeric, for Linux (x86-64 and ARM64, with CUDA 12/13 and multi-node support). GASNet-based (rather than UCX-based) conda packages are under the gex label. Windows is currently supported through WSL.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/26.01/.

Highlights

Added functionality

• Implement cupynumeric.pad.
• Implement cupynumeric.linalg.pinv (single-CPU/GPU only)
• Implement from_dlpack for exporting cuPyNumeric ndarray s through the DLPack interface
• Detect when an object being used to initialize a cuPyNumeric ndarray implements the DLPack interface, and use it if possible.

Bugfixes

• Ensure unimplemented stub functions always return cuPyNumeric ndarrays.

Known issues

• We are aware of hangs when using cuSolverMp-based APIs on 4+ Perlmutter nodes. This appears to be a cluster-specific issue, that we are investigating.
• We are aware of performance regressions with cupynumeric.einsum on Blackwell GPUs, starting to occur with cuBLAS 13.2. These are under investigation.

Full Changelog: v25.11.00...v26.01.00

v25.11.00

This is a beta release of cuPyNumeric.

Pip wheels are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/, for Linux (x86-64 and ARM64, with CUDA 12 and multi-node support) and macOS (for ARM64). Conda packages are available at https://anaconda.org/legate/cupynumeric, for Linux (x86-64 and ARM64, with CUDA 12/13 and multi-node support). GASNet-based (rather than UCX-based) conda packages are under the gex label. Windows is currently supported through WSL.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.11/.

Highlights

Support matrix changes

• Start distributing conda packages for CUDA 13.
• Port to cuSolverMp 0.7 (now the new required minimum).
• Validate cuPyNumeric on DGX Spark.

Note that currently the pip wheels do not include CUDA 13 support, nor cuSolverMp support (linear solve / matrix decomposition APIs are constrained to single-GPU execution when using the wheels).

Added functionality

• cupynumeric.histogram2d and cupynumeric.histogramdd
• cupynumeric.lexsort
• cupynumeric.isin
• Multi-GPU & multi-node implementation of QR factorization, based on cuSolverMp

Performance improvements

• Accelerate axis-wise reductions on GPUs by combining multiple kernel invocations into one.
• Parallelize specialized implementation for cupynumeric.take, and use it in more cases, including cupynumeric.take_along_axis.

UX improvements

• I/O functions (e.g. hdf5 to_file) and memory offloading (e.g. offload_to) functions from Legate now accept cuPyNumeric ndarrays directly.

Known issues

• We are aware of hangs when using cuSolverMp-based APIs on 4+ Perlmutter nodes. This appears to be a cluster-specific issue, that we are investigating.
• We are aware of hangs when using UCX 1.19 with the CUDA 13 conda packages. These are typically accompanied by an error message like this:

  ib_md.c:287  UCX  ERROR ibv_reg_mr(address=(nil), length=134217728, access=0xf) failed: Bad address
  ucp_mm.c:76   UCX  ERROR failed to register address (nil) (cuda) length 134217728 on md[6]=mlx5_0: Input/output error (md supports: host|cuda)
  

  We are investigating a proper fix. For the time being, setting UCX_MEMTYPE_CACHE=no in the environment appears to resolve the hang, at the cost of potentially decreasing UCX performance.

Full Changelog: v25.10.00...v25.11.00

v25.10.00

This is a beta release of cuPyNumeric.

Pip wheels are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/, for Linux (x86-64 and ARM64, with CUDA and multi-node support) and macOS (for ARM64). Conda packages are available at https://anaconda.org/legate/cupynumeric, for Linux (x86-64 and ARM64, with CUDA and multi-node support). GASNet-based (rather than UCX-based) conda packages are under the gex label. Windows is currently supported through WSL.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.10/.

Highlights

Added functionality

• Implement cupynumeric.in1d.
• Add DLPack import/export support to cuPyNumeric ndarrays.
• Allow batched input for cupynumeric.linalg.solve.

Performance improvements

• Optimized implementation for the special axis= case of cupynumeric.take.
• Improve heuristics for choosing between batched and unbatched matrix multiplication.
• Improved implementation of cupynumeric.nonzero that uses no additional scratch space.
• Identify special cases of advanced indexing that can be executed faster using cupynumeric.einsum.

Documentation / profiling

• Add a tutorial on using Legate Tasks to extend cuPyNumeric.
• Add a user warning when an operation (e.g. printing to the console) causes a sharded array to be gathered onto a single memory.
• Add sub-boxes to the Legate profiler, showing how long the Python interpreter spends inside cuPyNumeric API calls.

Breaking changes

• Move nightly conda packages to a dedicated channel, -c legate-nightly.

Known issues

• We are aware of hangs occurring under certain platforms and UCC configurations, when using cuSolverMp-backed multi-GPU operations (Cholesky factorization and linear solve). We expect these to be fixed by the 25.11 release, that updates to cuSolverMp 0.7.

Full Changelog: v25.08.00...v25.10.00

v25.08.00

This is a beta release of cuPyNumeric.

Pip wheels are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/, for Linux (x86-64 and ARM64, with CUDA and multi-node support) and macOS (for ARM64). Conda packages are available at https://anaconda.org/legate/cupynumeric, for Linux (x86-64 and ARM64, with CUDA and multi-node support). GASNet-based (rather than UCX-based) conda packages are under the gex label. Windows is currently supported through WSL.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.08/.

New features

Added functionality

• Multi-node multi-GPU capable SVD, specialized for tall-skinny matrices
• cupynumeric.cross
• cupynumeric.insert
• cupynumeric.logspace
• cupynumeric.real_if_close
• cupynumeric.roots
• cupynumeric.ravel_multi_index
• cupynumeric.copyto
• cupynumeric.diagflat
• cupynumeric.delete
• cupynumeric.nan_to_num
• Support multi-axis reductions

Performance Improvements

• Improve robustness & speed of cupynumeric.sort, by combining allocations where possible, and adding synchronization barriers around NCCL collectives.
• Remove some extraneous blocking that was only necessary to match the behavior of NumPy 1.x.
• Improve performance of NumPy fallback, in particular removing extraneous array copies, and adding special cases for quick fallback to functions such as cupynumeric.concatenate.

Miscellaneous

• Unify all environment variables that control cuPyNumeric's NumPy fallback heuristics, to a single one, CUPYNUMERIC_MAX_EAGER_VOLUME.
• Allow any available BLAS implementation to be used in a source build.

Full Changelog: v25.07.00...v25.08.00

v25.07.00

This is a beta release of cuPyNumeric.

Pip wheels are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/, for Linux (x86-64 and ARM64, with CUDA and multi-node support) and macOS (for ARM64). Conda packages are available at https://anaconda.org/legate/cupynumeric, for Linux (x86-64 and ARM64, with CUDA and multi-node support). GASNet-based (rather than UCX-based) conda packages are under the gex label. Windows is currently supported through WSL.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.07/.

New features

Added functionality

• Multi-node multi-GPU capable cupynumeric.linalg.solve and cupynumeric.linalg.cholesky, backed by cuSolverMp.
• Single-GPU cupynumeric.linalg.eigh/eigvalsh, backed by cuSolver.
• cupynumeric.round

Support matrix changes

• macOS wheels are now available on PyPI.
• Add support for Blackwell CUDA architecture and MNNVL.
• Drop support for Python 3.10 and add support for Python 3.13.
• Remove NumPy 1.X restriction from packages (now compatible with NumPy 2.X).

Tuning

• Add an optional "doctor" mode, that will detect some common anti-patterns causing bad performance. Enable with CUPYNUMERIC_DOCTOR=1, see https://docs.nvidia.com/cupynumeric/25.07/api/settings.html#doctor for more information.

Documentation

• A basic cuPyNumeric tutorial is available, see https://docs.nvidia.com/cupynumeric/25.07/user/tutorial.html.
• Start publishing nightly doc builds to https://nv-legate.github.io/cupynumeric.

Full Changelog: v25.03.02...v25.07.00

Known issues

• Multi-node runs can occasionally segfault at exit. This issue is under investigation. Preliminary investigation suggests that the issue depends on the ordering between cuPyNumeric and OpenBLAS teardown. There is no impact to the correctness of the computation and subsequent GPU usage.
• If the user explicitly forces multi-GPU execution of a sorting operation on very small arrays (about as many elements as the number of GPUs) this can result in CUDA errors. In normal conditions cuPyNumeric would not be GPU-accelerating operations of this size. A fix for this issue is in development and will be made available in an upcoming nightly build.

v25.03.02

This is a beta release of cuPyNumeric.

Linux x86 and ARM builds for Python 3.10 - 3.12 are available on PyPI at https://pypi.org/project/nvidia-cupynumeric/, and as conda packages at https://anaconda.org/legate/cupynumeric.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.03/.

New features

PIP install support

With this release, Linux x86 and ARM builds of cuPyNumeric are available for Python 3.10 - 3.12 as Python wheels on PyPI in addition to conda.

• cuPyNumeric can be installed with:

  pip install nvidia-cupynumeric
  

  See https://docs.nvidia.com/cupynumeric/25.03/installation.html#installing-pypi-packages for further instructions.
• These wheels support multi-node execution through UCX.
  See https://docs.nvidia.com/legate/25.03/networking-wheels.html for more details.

v25.03.00

This is a beta release of cuPyNumeric.

Linux x86 and ARM conda packages are available for this release at https://anaconda.org/legate/cupynumeric.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.03/.

New features

Licensing

• With this release the Legate framework, on which cuPyNumeric is based, becomes open-source, under the Apache-2.0 license. This makes the entire cuPyNumeric stack (anything above the CUDA library level) open-source.

Added functionality

• Matrix exponential: cupynumeric.linalg.expm
• Batched eigendecomposition: cupynumeric.linalg.eigvals & cupynumeric.linalg.eig

Performance improvements

• No longer doing unnecessary streaming when running matrix multiplication on a single processor/GPU.

UX improvements

• Add thelegate.core.ProfileRange Python context manager, to annotate sub-spans within a larger task span on the profiler visualization.
• Add the local_task_array helper function, that can be used in Python tasks to create a view over a Store/Array argument, using a NumPy or CuPy array as appropriate based on the type of memory where the data is located.

Documentation improvements

• Add a user guide chapter on accelerating multi-GPU HDF5 workloads.

Known issues

• We are aware of possible performance regressions when using UCX 1.18. We are temporarily restricting our packages to UCX <= 1.17 while we investigate this.

v25.01.00

This is a beta release of cuPyNumeric.

Linux x86 and ARM conda packages are available at https://anaconda.org/legate/cupynumeric.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/25.01/.

New features

Added functionality

• Add the method parameter to cupynumeric.convolve.
• Increase the maximum array dimension from 4 to 6.
• Experimental support for NumPy 2.0 (not reflected in the package constraints yet).

Memory management enhancements

• Updates to take advantage of the deferred-eager pool unification in Legate. This change has the potential to increase the effective available memory capacity by up to 100% for many usecases. It also removes the need for the user to adjust the --eager-alloc-percentage.
• Add the offload_to() API, that allows a user to offload an array to a particular memory kind, such that any copies in other memories are discarded. This can be useful e.g. to evict an array from GPU memory onto system memory, freeing up space for subsequent GPU tasks.

I/O improvements

• Use cuFile to accelerate HDF5 reads on the GPU.
• Add support for reading "binary" HDF5 datasets (in particular useful for reading boolean-type datasets).

UX Improvements

• Consider NUMA node topology when allocating CPU cores and memory during automatic machine configuration.
• Add environment variable LEGATE_LIMIT_STDOUT, to only print out the output from one of the copies of the top-level program in a multi-process execution.
• Remove an extraneous warning about __buffer__ being unimplemented.

Deprecations

• Drop support for the Maxwell GPU architecture. cuPyNumeric now requires at least Pascal (sm_60).

v24.11.02

This is a patch release of cuPyNumeric.

Linux x86 and ARM conda packages are available at https://anaconda.org/legate/cupynumeric.

Documentation for this release can be found at https://docs.nvidia.com/cupynumeric/24.11/.

Packaging Changes

• Update for Legate v24.11.01