You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Support managed memory advise, prefetch, and discard-prefetch #1332
Status update: the blocker (#718ManagedMemoryResource) was closed on 2025-12-16, so this issue itself is unblocked.
Work is happening in PR #1775 by @rparolin, but it has been inactive since 2026-03-19 and has a CHANGES_REQUESTED review from @leofang (2026-03-25). Full triage posted on the PR: #1775
Two design decisions need to be resolved before the PR can merge:
Free functions under cuda.core.managed_memory vs. a ManagedBuffer(Buffer) subclass with property-style API (Andy-Jost's proposal).
Whether single-range advise/prefetch/discard_prefetch lands together with the batched variants (Support batched memory movement #1333) or separately.
Plus mechanical issues (rebase against main, cydriver / IF/ELSE, options dataclass, module-placement nit).
Given v1.0.0 is due 2026-05-07 (14 days) and the design is not yet settled, this may need to slip to the post-1.0 milestone unless the design decisions close within a few days.
I can pick this issue up after the v1.0.0 release. We have agreement all around on using property-like access and richer location specifiers such as Host(numa_id=1).
Activity
leofang commentedon Apr 7, 2026
Pushing this out after discussing with Andy
cpcloud commentedon Apr 23, 2026
Status update: the blocker (#718
ManagedMemoryResource) was closed on 2025-12-16, so this issue itself is unblocked.Work is happening in PR #1775 by @rparolin, but it has been inactive since 2026-03-19 and has a
CHANGES_REQUESTEDreview from @leofang (2026-03-25). Full triage posted on the PR: #1775Two design decisions need to be resolved before the PR can merge:
cuda.core.managed_memoryvs. aManagedBuffer(Buffer)subclass with property-style API (Andy-Jost's proposal).advise/prefetch/discard_prefetchlands together with the batched variants (Support batched memory movement #1333) or separately.Plus mechanical issues (rebase against main,
cydriver/IF/ELSE, options dataclass, module-placement nit).Given v1.0.0 is due 2026-05-07 (14 days) and the design is not yet settled, this may need to slip to the post-1.0 milestone unless the design decisions close within a few days.
Andy-Jost commentedon Apr 27, 2026
I can pick this issue up after the v1.0.0 release. We have agreement all around on using property-like access and richer location specifiers such as
Host(numa_id=1).