LocalizedAllocator

class torch.cuda.memory.LocalizedAllocator(locality_domain_id, *, device=None)

Allocate physical CUDA memory on one locality domain.

Use MemPool(allocator=allocator.allocator()) to cache and suballocate this memory. Requires CUDA driver and cuda.bindings 13.4+ and a multi-domain GPU. Construction initializes CUDA to validate the actual device topology. Kernel execution is not localized by this allocator.

Parameters:

• locality_domain_id (int) – Locality domain to allocate from, excluding bool.

• device (torch.device or int, optional) – Owning CUDA device. Defaults to the current device.

Warning

Concurrent allocator activity is unsupported. Allocation, freeing, and allocator-state queries (including MemPool.use_count()) from other threads must not overlap use of this allocator. Native allocator callbacks acquire the GIL while holding the allocator mutex, which can deadlock with a thread holding the GIL while waiting for that mutex.

Note

Use this allocator only on its owning device; access from other devices is not supported. Callback owners are retained for the process lifetime, so tensors may outlive the Python allocator and pool objects. One owner and primary-context reference per device/domain are deliberately retained, including during interpreter shutdown. Releasing physical memory waits for its allocation stream. Freeing is best-effort: failures are logged, not raised. Memory that cannot safely be unmapped is left for the driver to reclaim at exit.

property device_id: int

Return the CUDA device index associated with this allocator.

property locality_domain_id: int

Return the locality domain ID associated with this allocator.