EPIC: VMM feature completeness #2980

Activity

1. leofang

   added

   P0High priority - Must do!High priority - Must do!

   featureNew feature or requestNew feature or request

   cuda.coreEverything related to the cuda.core moduleEverything related to the cuda.core module

   blockedThis task is currently blocked by other tasksThis task is currently blocked by other tasks

   on Sep 30, 2026
2. leofang commented on Sep 30, 2026

   

   leofang

   on Sep 30, 2026

   MemberAuthor

   More actions

   @Andy-Jost to investigate remaining gaps, break down tasks to do, and fan out to the team as needed.
3. leofang

   assigned

   Andy-Jost

   on Sep 30, 2026
4. leofang commented on Sep 30, 2026

   

   leofang

   on Sep 30, 2026

   MemberAuthor

   More actions

   Tentatively set to v1.3.0. Andy to make the final call.
5. leofang

   added this to the cuda.core 1.3.0 milestone on Sep 30, 2026
6. Andy-Jost

   removed

   blockedThis task is currently blocked by other tasksThis task is currently blocked by other tasks

   on Oct 5, 2026
7. Andy-Jost

   added a commit that references this issue on Oct 14, 2026

   cuda.core: share VirtualMemoryResource buffers between processes (#3028)

   ...

   1f41281
8. Andy-Jost commented on Oct 23, 2026

   

   Andy-Jost

   on Oct 23, 2026

   Contributor

   More actions

   Status update. The target for 1.3.0 is VirtualMemoryResource with graph support, IPC, and fabric handles.

   • Done: the handle-layer rewrite (fix(cuda.core): move VirtualMemoryResource onto the _rt handle layer #2917) made VMM buffers safe in graphs; a captured copy keeps the range mapped until the launch completes. cuda.core: share VirtualMemoryResource buffers between processes #3028 added sharing of VMM buffers between processes (POSIX file descriptors, Linux), with forkserver coverage in test(cuda.core): cover the forkserver start method for VMM IPC buffers #3124.
   • Follow-ups from the IPC review: pickled resources and devices now travel by UUID instead of ordinal (fix(cuda.core): pickle memory-resource devices by UUID for IPC #3126 merged, fix(cuda.core): pickle Device by UUID #3198 in review), and fd pressure for buffers with many chunks is tracked in cuda.core: transfer VMM IPC allocations one at a time #3110.
   • Remaining: fabric handles, the last item on this epic. A second descriptor class carries the 64-byte handles, with export and import, tests, and docs. Fabric needs an IMEX channel, which no CI runner has, so the tests skip in CI and validation runs on internal Blackwell systems. This lands before the 1.3.0 cut.