cuda.core 1.3.1 Release Notes#

Highlights#

  • New minimal-sysctk12 and minimal-sysctk13 extras install only cuda-bindings at or above this release’s floor for that CUDA major, without the CUDA Toolkit wheels that cu12 and cu13 add. The cu12 and cu13 extras include them. Packages that depend on cuda-core should request one of these extras, so that pip and uv see the floor; with a plain cuda-core requirement, an older cuda-bindings is installed and fails at import. See cuda-bindings Version Requirements. (#3459)

  • cuda-core 1.3.0 is yanked from PyPI. It declared its floors only in the cu12 and cu13 extras, so a resolver that cannot use 1.3.1 could fall back to it next to an older cuda-bindings.

Fixes and enhancements#

  • Buffer.from_ipc_descriptor() no longer unmaps a newly imported memory pool buffer when another thread releases the same descriptor at the same time. The import now waits for the pending release to finish. (#3450)

  • PinnedMemoryResource construction without options no longer invalidates a CUDA graph capture in progress on another thread. cuda.core 1.3.0 introduced a cuMemPoolSetAccess call in the constructor (#3405); the driver refuses that call, and invalidates any concurrent global-mode capture, unless it runs in relaxed capture mode. The fix applies the same relaxed-mode wrapper already used for pool attribute reads and writes and trim(). Captures started with the default begin_building() mode are unaffected; global mode is the default in several external libraries (for example torch.cuda.graph). (#3451)