gpo.zugaina.org

Search Portage & Overlays:

dev-python/cuda-core

cuda.core: pythonic CUDA module

Screenshots

  • cuda-core-1.2.0
    ~amd64
    python_targets_python3_12 python_targets_python3_13 python_targets_python3_14 debug

    View      Download      Browse     License: Apache-2.0   
    Overlay: stuff
  • cuda-core-1.1.1
    ~amd64
    python_targets_python3_12 python_targets_python3_13 python_targets_python3_14 debug

    View      Download      Browse     License: Apache-2.0   
    Overlay: stuff

ChangeLog

commit 738f8a32ba3f4a0ba01832a2ce1b0f80ab9d7e7b
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 03:59:28 2026 +0200

dev-python/cuda-core: trim 1.1.1 comment prose

commit e854c4cdfd7e9705a3833e9a1c4edf567310d0b1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:52:38 2026 +0200

dev-python/cuda-core: trim 1.2.0 comment prose

commit 4f0ddf174ac99618b9dbfa525389bcd0d97352be
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Sep 4 05:11:25 2026 +0200

dev-python/cuda-core: re-pin 1.2.0 distfile

The archive GitHub generates for the cuda-core-v1.2.0 tag changed size,
so the recorded hashes no longer match what is served and the fetch
aborts. The current archive's pax header carries the commit that the
tag resolves to, and the 1.1.1 archive still matches its pin, so this
looks like an encoding change rather than a moved tag -- but the
superseded archive was not available here, so that reading is inference
rather than verification. Re-pin to the bytes that are actually
fetchable and record how far the check went, so a later audit can tell
this apart from real upstream tampering instead of trusting the note.

commit 62cbad92c47f18d1dcc8d3b2a4b09df54f57d454
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 3 14:43:04 2026 +0200

dev-python/cuda-core: drop 1.1.0

Down to the last two. Consumers use floors and a <2 cap
(>=0.3.2, >=0.5.1, <2), all satisfied by 1.1.1 and 1.2.0.

commit 513bf99bb08ee7bf13614ddcfdf169c2e1869a8b
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 3 14:42:51 2026 +0200

dev-python/cuda-core: add 1.2.0

Dependency set unchanged, read from the monorepo tarball's
cuda_core/pyproject.toml: cuda-pathfinder>=1.4.2, numpy, and the
python_version<'3.11' backports.strenum that PYTHON_COMPAT
python3_ makes moot. The cuda-core-v<PV> tag prefix still holds,
so SRC_URI and S need no change.

Build-checked for real rather than deferred to host gating. The first
attempt failed with ModuleNotFoundError on cuda.pathfinder because
cuda-pathfinder is a BDEPEND that simply was not merged here; installing
the four build deps -- cuda-pathfinder-1.8.1, pyclibrary-0.3.0,
cuda-bindings-13.3.1-r2 and cython downgraded to 3.2.9 -- made the check
real, and Cython then compiled the .pyx sources (_device, _dlpack,
_device_resources, ...) as it should.

The cython downgrade is not incidental: both this package and
cuda-bindings cap <dev-python/cython-3.3 to match upstream's
build-system pin, so 3.2.x is the band these are meant to build in.

commit 552e90f3a817e8c4b69b92881ee3c51f7f7b0d42
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Aug 13 19:16:12 2026 +0200

dev-python/cuda-core: drop 1.0.1-r1

The retained newer releases provide the required rollback coverage without carrying
this superseded pre-1.1 revision.

commit c71869693eac28f0ccc9a102dd72037cd05f529f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jul 30 11:09:17 2026 +0200

dev-python/cuda-core: add 1.1.1

Build-system requires (setuptools, setuptools-scm, Cython <3.3,
cuda-pathfinder >=1.5) and runtime dependencies are unchanged from
1.1.0 -- upstream pyproject matches BDEPEND/RDEPEND -- so the ebuild is
copied verbatim.

commit cffc844e38c3890d6869830b9a85d7cd4095f0e8
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Jul 10 14:08:08 2026 +0200

dev-python/cuda-core: add 1.1.0

commit 8f8ca5d14ad8bf9e63cd3e43cb6b75a7b2ef1434
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 10 15:24:22 2026 +0200

dev-python/cuda-core: fix v-prefixed __version__ (-r1)

SETUPTOOLS_SCM_PRETEND_VERSION is used verbatim and bypasses tag_regex;
packaging normalises the v away in dist metadata but it leaks into
cuda.core.__version__ as "v1.0.1". Same defect as the cuda-bindings fix.
Use the literal $.

commit 3fc9c564bbda3f8e8ec26d5c71d0d315b2f2dc27
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 30 15:32:48 2026 +0200

dev-python/cuda-core: new package, 1.0.1

NVIDIA's pythonic CUDA module from the NVIDIA/cuda-python monorepo
(cuda-core-v<PV> tag prefix, like cuda-pathfinder). Apache-2.0 source,
runtime hits /opt/cuda headers via cuda.pathfinder. Required as a
runtime dep by dev-python/cuda-python ~=1.0.0 since 13.3.0; the
overlay was silently missing the RDEPEND entry until this landed.

cuda-bindings is added at both DEPEND and BDEPEND because cuda-core's
.pyx files do `from cuda.bindings cimport cydriver`, which needs the
installed package's .pxd headers — upstream pyproject only declares
this dep under the cu12/cu13 runtime extras but the build can't
proceed without it.