Skip to content

Commit 98c7671

Browse files
committed
docs(cuda.core): move managed-memory notes from 1.0.0 to 1.1.0
Per PR NVIDIA#1775 review feedback: this PR targets the cuda.core v1.1.0 milestone, so the new-feature entries belong in a new 1.1.0-notes.rst rather than being appended to the already-released 1.0.0 notes. The release index uses a glob pattern (release/*-notes), so the new file is picked up automatically.
1 parent 0faecfa commit 98c7671

2 files changed

Lines changed: 45 additions & 34 deletions

File tree

cuda_core/docs/source/release/1.0.0-notes.rst

Lines changed: 0 additions & 34 deletions
Original file line numberDiff line numberDiff line change
@@ -30,40 +30,6 @@ Highlights
3030
New features
3131
------------
3232

33-
- Added :class:`Host` as the symmetric counterpart of :class:`Device` for
34-
expressing managed-memory locations: ``Host()`` (any host),
35-
``Host(numa_id=N)`` (specific NUMA node), and ``Host.numa_current()``
36-
(calling thread's NUMA node).
37-
38-
- Added :class:`ManagedBuffer`, a :class:`Buffer` subclass returned by
39-
:meth:`ManagedMemoryResource.allocate` that exposes a property-style
40-
advice API:
41-
42-
- ``buf.read_mostly`` (bool) — driver-backed get/set.
43-
- ``buf.preferred_location`` (:class:`Device` | :class:`Host` | None) —
44-
driver-backed get/set; assigning ``None`` unsets.
45-
- ``buf.accessed_by`` — a live, set-like view; ``add()`` / ``discard()``
46-
issue advice, iteration queries the driver.
47-
- ``buf.prefetch(location, *, stream)``, ``buf.discard(*, stream)``,
48-
``buf.discard_prefetch(location, *, stream)`` — instance methods that
49-
delegate to the matching free functions.
50-
51-
Use :meth:`ManagedBuffer.from_handle` to wrap an existing managed-memory
52-
pointer.
53-
54-
- Added batched managed-memory range operations to :mod:`cuda.core.utils`
55-
(CUDA 13+): :func:`~utils.prefetch_batch`, :func:`~utils.discard_batch`,
56-
and :func:`~utils.discard_prefetch_batch`. Each takes a sequence of
57-
managed :class:`Buffer` instances and dispatches to the corresponding
58-
``cuMem*BatchAsync`` driver entry point, addressing the managed-memory
59-
portion of #1333. Single-buffer operations are exposed as instance
60-
methods on :class:`ManagedBuffer` (:meth:`~ManagedBuffer.prefetch`,
61-
:meth:`~ManagedBuffer.discard`, :meth:`~ManagedBuffer.discard_prefetch`)
62-
and as property setters (:attr:`~ManagedBuffer.read_mostly`,
63-
:attr:`~ManagedBuffer.preferred_location`,
64-
:attr:`~ManagedBuffer.accessed_by`). Locations are expressed via
65-
:class:`Device` or :class:`Host`.
66-
6733
- :meth:`Program.compile` now accepts an optional ``cache=`` keyword argument
6834
for avoiding recompilation of identical source + options + target. Two
6935
concrete implementations of the :class:`~utils.ProgramCacheResource` ABC are
Lines changed: 45 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,45 @@
1+
.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2+
.. SPDX-License-Identifier: Apache-2.0
3+
4+
.. currentmodule:: cuda.core
5+
6+
``cuda.core`` 1.1.0 Release Notes
7+
=================================
8+
9+
10+
New features
11+
------------
12+
13+
- Added :class:`Host` as the symmetric counterpart of :class:`Device` for
14+
expressing managed-memory locations: ``Host()`` (any host),
15+
``Host(numa_id=N)`` (specific NUMA node), and ``Host.numa_current()``
16+
(calling thread's NUMA node).
17+
18+
- Added :class:`ManagedBuffer`, a :class:`Buffer` subclass returned by
19+
:meth:`ManagedMemoryResource.allocate` that exposes a property-style
20+
advice API:
21+
22+
- ``buf.read_mostly`` (bool) — driver-backed get/set.
23+
- ``buf.preferred_location`` (:class:`Device` | :class:`Host` | None) —
24+
driver-backed get/set; assigning ``None`` unsets.
25+
- ``buf.accessed_by`` — a live, set-like view; ``add()`` / ``discard()``
26+
issue advice, iteration queries the driver.
27+
- ``buf.prefetch(location, *, stream)``, ``buf.discard(*, stream)``,
28+
``buf.discard_prefetch(location, *, stream)`` — instance methods that
29+
delegate to the matching free functions.
30+
31+
Use :meth:`ManagedBuffer.from_handle` to wrap an existing managed-memory
32+
pointer.
33+
34+
- Added batched managed-memory range operations to :mod:`cuda.core.utils`
35+
(CUDA 13+): :func:`~utils.prefetch_batch`, :func:`~utils.discard_batch`,
36+
and :func:`~utils.discard_prefetch_batch`. Each takes a sequence of
37+
managed :class:`Buffer` instances and dispatches to the corresponding
38+
``cuMem*BatchAsync`` driver entry point, addressing the managed-memory
39+
portion of #1333. Single-buffer operations are exposed as instance
40+
methods on :class:`ManagedBuffer` (:meth:`~ManagedBuffer.prefetch`,
41+
:meth:`~ManagedBuffer.discard`, :meth:`~ManagedBuffer.discard_prefetch`)
42+
and as property setters (:attr:`~ManagedBuffer.read_mostly`,
43+
:attr:`~ManagedBuffer.preferred_location`,
44+
:attr:`~ManagedBuffer.accessed_by`). Locations are expressed via
45+
:class:`Device` or :class:`Host`.

0 commit comments

Comments
 (0)