Repository navigation
cuda.core: VirtualMemoryResourceOptions accepts values the driver rejects, and modify_allocation mishandles edge inputs #2910
Description
Validation gaps found while analyzing #2882, each a small change: (1) addr_align larger than the granularity makes the adjacent-reservation probe fail with a misaligned hint, which is treated as fatal instead of falling back to the slow path; (2) new_size < buf.size is silently treated as an access-only update; (3) allocation_type=MANAGED is exposed but current drivers reject it in cuMemCreate; (4) modify_allocation guards a user-reachable condition with assert; (5) allocate(0) reaches the driver and fails while sibling resources return an empty Buffer; (6) the default handle_type=POSIX_FD fails on native Windows WDDM devices and on location_type="host"; (7) for host locations device_id is validated and ignored, HOST_NUMA_CURRENT can never allocate, and a host resource with peers reports is_device_accessible=False. Related: #2694.
Activity
bishal0922 commentedon Sep 30, 2026
I'd like to pick up items (4) and (5) from this list: raise a proper exception instead of the
assertinmodify_allocation, and haveallocate(0)return an emptyBufferlike the sibling resources. Happy to keep those to one small PR, if that split works for you. Could I be assigned?fix(cuda.core): move VirtualMemoryResource onto the _rt handle layer (#…
Andy-Jost commentedon Oct 2, 2026
Thanks for the offer, @bishal0922. #2917 has since merged and took care of items 1, 2, 4 and 5 and the host half of item 6, so the
assertandallocate(0)are already handled.Andy-Jost commentedon Oct 22, 2026
Status after #2917: items 1, 2, 4 and 5 and the host half of item 6 are fixed. Still open:
allocation_type="managed"is still exposed on CUDA 13 builds, and current drivers reject it incuMemCreate.handle_type="posix_fd"still fails on native Windows WDDM devices.device_idis validated and then ignored,"host_numa_current"can never allocate, and a host resource withpeersreportsis_device_accessible=False.Item 7 overlaps #2694. I plan to fix the two together.
Andy-Jost commentedon Oct 23, 2026
To be clear about scope: #3149 by @somtri covers the
host_numaandhost_numa_currenthalf of item 7. The other item-7 gaps (the ignoreddevice_idfor host resources, andpeerson a host resource) stay here for a follow-up after #3149 lands.