Skip to content

Commit d9b1aa4

Browse files
committed
cuda.core: correct the 1.3.0 note for pool attributes, move getter fix to Fixes
1 parent 7846da3 commit d9b1aa4

1 file changed

Lines changed: 15 additions & 8 deletions

File tree

‎cuda_core/docs/source/release/1.3.0-notes.rst‎

Lines changed: 15 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -78,14 +78,14 @@ New features
7878

7979
- The memory pool attributes of :class:`DeviceMemoryResource`,
8080
:class:`PinnedMemoryResource` and :class:`ManagedMemoryResource` can now be
81-
set: ``release_threshold`` and the three ``reuse_*`` policies take any
82-
value, and ``reserved_mem_high`` and ``used_mem_high`` can be set to zero to
83-
reset the watermark, as for :class:`GraphMemoryResource`. The new
84-
``trim(min_bytes_to_keep=0)`` method returns unused pool memory to the OS.
85-
Reading or setting a pool attribute, and ``trim()``, no longer fail with
86-
``CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED`` and invalidate a stream capture
87-
that is in progress in global mode on any thread, or in thread-local mode
88-
on the calling thread.
81+
set: ``release_threshold`` takes a non-negative integer, the three
82+
``reuse_*`` policies take a bool, and ``reserved_mem_high`` and
83+
``used_mem_high`` can be set to zero to reset the watermark, as for
84+
:class:`GraphMemoryResource`. The new :meth:`DeviceMemoryResource.trim`
85+
method, also on the other two resources, returns unused pool memory to the
86+
OS. Setting an attribute and ``trim()`` take effect immediately, even during
87+
a stream capture, and leave the capture valid.
88+
(`#3261 <https://github.com/NVIDIA/cuda-python/issues/3261>`__)
8989

9090
- Added :class:`VirtualMemoryBuffer`, the :class:`Buffer` subclass that
9191
:class:`VirtualMemoryResource` returns. Closing it on a capturing stream
@@ -166,6 +166,13 @@ Fixes and enhancements
166166
new pool with options during such a capture is still refused by the driver.
167167
(`#2834 <https://github.com/NVIDIA/cuda-python/issues/2834>`__)
168168

169+
- Reading the :attr:`~DeviceMemoryResource.attributes` of a
170+
:class:`DeviceMemoryResource`, :class:`PinnedMemoryResource` or
171+
:class:`ManagedMemoryResource` no longer fails, or invalidates a stream
172+
capture, while the calling thread is inside a ``thread_local`` capture or
173+
any thread is inside a ``global`` one.
174+
(`#3236 <https://github.com/NVIDIA/cuda-python/pull/3236>`__)
175+
169176
- Ending, closing, or collecting a top-level :class:`~graph.GraphBuilder`
170177
whose capture was invalidated, or left unjoined by a
171178
:meth:`~graph.GraphBuilder.join` that failed partway, no longer crashes.

0 commit comments

Comments
 (0)