ciflow/trunk/201842: [symm_mem] Free reserved virtual address ranges on teardown
- PyTorch: 2736 events in the last 90 days
- PyTorch: 2693th Release in the last 90 days
- Previous: earlier the same day · ciflow/torchtitan/201769: Update
What happened
map_block() reserves a virtual address (VA) range with cuMemAddressReserve. No code frees this range. ~AllocationRef unmaps and releases the memory, but it does not call cuMemAddressFree. The multicast cleanup in MulticastSetup::release() has the same problem. Each alloc and free leaks one VA range of block_size. Each peer mapping and each multicast mapping from rendezvous also leaks one range. The physical memory i…
Summary assembled by rule from the sources below