cuda.bindings._v2.driver.checkpoint_operation_complete#

cuda.bindings._v2.driver.checkpoint_operation_complete(intptr_t handle)#

Complete a custom-storage checkpoint or restore operation.

After cuCheckpointProcessCheckpoint or cuCheckpointProcessRestore with custom storage, cuCheckpointOperationComplete should be called when the application has finished or enqueued (on a stream) the required copies (for checkpoint, GPU mappings to custom storage and for restore, custom storage into GPU mappings) so the driver can finish the operation. cuCheckpointOperationComplete will first synchronize on the streams returned by cuCheckpointProcessCheckpoint or cuCheckpointProcessRestore. The driver assumes that the application has completed copying the memory to or from custom storage once synchronization is done. None of the streams returned by cuCheckpointProcessCheckpoint or cuCheckpointProcessRestore should be in stream capture mode; any attempt to do so would result in CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED. Once the synchronization is done, cuCheckpointOperationComplete will unmap the GPU memory that was mapped on each GPU in a particular cuCheckpointProcessCheckpoint or cuCheckpointProcessRestore operation. cuCheckpointOperationComplete identifies the operation to be completed using handle. This handle is set by cuCheckpointProcessCheckpoint or cuCheckpointProcessRestore. When the call returns successfully, the target process will be in the CU_PROCESS_STATE_CHECKPOINTED state when completing a checkpoint operation or the CU_PROCESS_STATE_LOCKED state when completing a restore operation.

Parameters:: handle (intptr_t) – CUcheckpointCustomStorageInfo.handle set by a prior cuCheckpointProcessCheckpoint or cuCheckpointProcessRestore call.