From 10309dbadce2ff0bd51b40d6b379db39230366dd Mon Sep 17 00:00:00 2001 From: Leo Fang Date: Wed, 2 Sep 2026 13:21:32 -0700 Subject: [PATCH 1/2] Prepare cuda.core v1.2.0 release Fill remaining user-visible entries in the 1.2.0 release notes and add 1.1.1 and 1.2.0 to the docs version switcher. Version is derived from the git tag via setuptools-scm, so no pyproject.toml bump is needed. New features documented: - ProgramOptions.use_bundled_headers for NVRTC 13.3+ (#2753 closes #2363) - cuda-gdb source display for JIT-compiled kernels (#2678, #2679) Fixes and enhancements documented: - PinnedMemoryResource host-pool validation at construction (#2487) - is_host_accessible for NUMA-located VMM (#2503) - Graph predecessor/successor query truncation on large graphs (#2587) - Per-domain NVML clock queries treated as independently optional (#2651) - Temperature threshold checks forward-compat with unknown archs (#2488) - import cuda.core non-integer opt-out crash (#2535) - Frozen CUresult explanation table refreshed for CUDA 13.3 (#2383) -- Leo's bot --- cuda_core/docs/nv-versions.json | 8 +++ cuda_core/docs/source/release/1.2.0-notes.rst | 56 +++++++++++++++++++ 2 files changed, 64 insertions(+) diff --git a/cuda_core/docs/nv-versions.json b/cuda_core/docs/nv-versions.json index a81563a3bfa..3bcd163e39b 100644 --- a/cuda_core/docs/nv-versions.json +++ b/cuda_core/docs/nv-versions.json @@ -3,6 +3,14 @@ "version": "latest", "url": "https://nvidia.github.io/cuda-python/cuda-core/latest/" }, + { + "version": "1.2.0", + "url": "https://nvidia.github.io/cuda-python/cuda-core/1.2.0/" + }, + { + "version": "1.1.1", + "url": "https://nvidia.github.io/cuda-python/cuda-core/1.1.1/" + }, { "version": "1.1.0", "url": "https://nvidia.github.io/cuda-python/cuda-core/1.1.0/" diff --git a/cuda_core/docs/source/release/1.2.0-notes.rst b/cuda_core/docs/source/release/1.2.0-notes.rst index f96a205d1e8..7bd25efb474 100644 --- a/cuda_core/docs/source/release/1.2.0-notes.rst +++ b/cuda_core/docs/source/release/1.2.0-notes.rst @@ -37,6 +37,22 @@ New features (`#2456 `__, `#1334 `__) +- Added :attr:`ProgramOptions.use_bundled_headers`, which lets NVRTC + resolve the CUDA and CCCL headers from the toolkit bundled with NVRTC + itself, installed into a per-user cache directory, instead of requiring + a full CUDA Toolkit installation on the compile host. NVRTC backend + only; requires NVRTC 13.3 or newer. + (`#2753 `__, + closes `#2363 `__) + +- When a :class:`Program` is compiled with ``debug`` or ``lineinfo``, + the NVRTC input source is now materialized as a temporary ``.cu`` file + so ``cuda-gdb`` can list the original source while stepping through + JIT-compiled kernels. ``#include "..."`` search still resolves against + the original source directory. + (`#2678 `__, + `#2679 `__) + Fixes and enhancements ---------------------- @@ -196,6 +212,46 @@ Fixes and enhancements from being treated as CUDA devices by :meth:`Device.get_all_devices`, examples, and tests. +- :class:`PinnedMemoryResource` now rejects unsupported host memory pools + at construction with ``RuntimeError``, instead of letting a later + allocation or copy fail with ``CUDA_ERROR_INVALID_VALUE``. + (`#2487 `__) + +- :attr:`VirtualMemoryResource.is_host_accessible`, and by extension + :attr:`Buffer.is_host_accessible`, now correctly return ``True`` for a + resource configured with ``location_type="host_numa"`` or + ``"host_numa_current"``. Previously both properties reported ``False`` + on those NUMA-located variants. + (`#2503 `__) + +- Graph predecessor and successor queries no longer truncate their results + on large graphs. + (`#2587 `__) + +- Per-domain clock queries in :mod:`cuda.core.system` treat each domain + (minimum, maximum, and current) as independently optional, so an + unsupported domain no longer fails the whole clock query for a device. + (`#2651 `__) + +- Temperature threshold checks in :mod:`cuda.core.system` are now + forward-compatible with GPU architectures newer than the generated + ``DeviceArch`` enum. An unrecognized architecture no longer raises + ``ValueError`` before the query runs. + (`#2488 `__) + +- ``import cuda.core`` no longer raises ``ValueError`` when + ``CUDA_CORE_DONT_FIX_TAB_COMPLETION`` is set to a non-integer value such + as an empty string, ``true``, or ``yes``. Any non-empty, non-integer + value is honored as an opt-out. See :doc:`environment_variables` for + details. + (`#2535 `__) + +- The frozen fallback ``CUresult`` explanation table, used when the + driver's ``cuGetErrorName`` / ``cuGetErrorString`` are unavailable, is + refreshed for CUDA 13.3 and now recognizes + ``CUDA_ERROR_GRAPH_RECAPTURE_FAILURE``. + (`#2383 `__) + Deprecation Notices ------------------- From 45bbf3002f3a8a0610b0d96fb55e2fb5d69ad9f8 Mon Sep 17 00:00:00 2001 From: Leo Fang Date: Wed, 2 Sep 2026 16:41:19 -0400 Subject: [PATCH 2/2] Update cuda_core/docs/source/release/1.2.0-notes.rst Co-authored-by: Michael Droettboom --- cuda_core/docs/source/release/1.2.0-notes.rst | 7 ------- 1 file changed, 7 deletions(-) diff --git a/cuda_core/docs/source/release/1.2.0-notes.rst b/cuda_core/docs/source/release/1.2.0-notes.rst index 7bd25efb474..dd1f3504e9a 100644 --- a/cuda_core/docs/source/release/1.2.0-notes.rst +++ b/cuda_core/docs/source/release/1.2.0-notes.rst @@ -239,13 +239,6 @@ Fixes and enhancements ``ValueError`` before the query runs. (`#2488 `__) -- ``import cuda.core`` no longer raises ``ValueError`` when - ``CUDA_CORE_DONT_FIX_TAB_COMPLETION`` is set to a non-integer value such - as an empty string, ``true``, or ``yes``. Any non-empty, non-integer - value is honored as an opt-out. See :doc:`environment_variables` for - details. - (`#2535 `__) - - The frozen fallback ``CUresult`` explanation table, used when the driver's ``cuGetErrorName`` / ``cuGetErrorString`` are unavailable, is refreshed for CUDA 13.3 and now recognizes