o h@sddlmZmZddlZddlmZddlmZee e ej fZ dgZ de dej fd d Zd ee e fdeej ej ffd d Zdee deej fddZGdddeZdS))OptionalUnionN) _TensorPipeRpcBackendOptionsBase) constantsTensorPipeRpcBackendOptionsdevicereturncCs*t|}|jdkrtd|jd|S)NcudazA`set_devices` expect a list of CUDA devices, but got device type .)torchrtype ValueError)rrq/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/torch/distributed/rpc/options.py _to_devices  r device_mapcCsji}i}|D]*\}}t|t|}}||vr*td|d||d||||<|||<q|S)Nz9`device_map` only supports 1-to-1 mapping, trying to map  and  to )itemsr rr)rfull_device_map reverse_mapkvrrr_to_device_maps" rdevicescCsttt|S)N)listmapr)rrrr_to_device_list*srcseZdZdZejejejddddddede de de e e e e e ffde ee d e ed e effd d Zd e de e e fffdd Zdee fddZZS)ra' The backend options for :class:`~torch.distributed.rpc.TensorPipeAgent`, derived from :class:`~torch.distributed.rpc.RpcBackendOptions`. Args: num_worker_threads (int, optional): The number of threads in the thread-pool used by :class:`~torch.distributed.rpc.TensorPipeAgent` to execute requests (default: 16). rpc_timeout (float, optional): The default timeout, in seconds, for RPC requests (default: 60 seconds). If the RPC has not completed in this timeframe, an exception indicating so will be raised. Callers can override this timeout for individual RPCs in :meth:`~torch.distributed.rpc.rpc_sync` and :meth:`~torch.distributed.rpc.rpc_async` if necessary. init_method (str, optional): The URL to initialize the distributed store used for rendezvous. It takes any value accepted for the same argument of :meth:`~torch.distributed.init_process_group` (default: ``env://``). device_maps (Dict[str, Dict], optional): Device placement mappings from this worker to the callee. Key is the callee worker name and value the dictionary (``Dict`` of ``int``, ``str``, or ``torch.device``) that maps this worker's devices to the callee worker's devices. (default: ``None``) devices (List[int, str, or ``torch.device``], optional): all local CUDA devices used by RPC agent. By Default, it will be initialized to all local devices from its own ``device_maps`` and corresponding devices from its peers' ``device_maps``. When processing CUDA RPC requests, the agent will properly synchronize CUDA streams for all devices in this ``List``. N)num_worker_threads rpc_timeout init_method device_mapsr _transports _channelsrr r!r"rr#r$c sN|durindd|D}|durgnt|} t||||||| dS)NcSsi|] \}}|t|qSr)r).0rrrrr ^sz8TensorPipeRpcBackendOptions.__init__..)rrsuper__init__) selfrr r!r"rr#r$full_device_mapsfull_device_list __class__rrr(Ps z$TensorPipeRpcBackendOptions.__init__torc szt|}tj}||vr4|D]#\}}|||vr3||||kr3td|d|d|||qt||dS)a0 Set device mapping between each RPC caller and callee pair. This function can be called multiple times to incrementally add device placement configurations. Args: to (str): Callee name. device_map (Dict of int, str, or torch.device): Device placement mappings from this worker to the callee. This map must be invertible. Example: >>> # xdoctest: +SKIP("distributed") >>> # both workers >>> def add(x, y): >>> print(x) # tensor([1., 1.], device='cuda:1') >>> return x + y, (x + y).to(2) >>> >>> # on worker 0 >>> options = TensorPipeRpcBackendOptions( >>> num_worker_threads=8, >>> device_maps={"worker1": {0: 1}} >>> # maps worker0's cuda:0 to worker1's cuda:1 >>> ) >>> options.set_device_map("worker1", {1: 2}) >>> # maps worker0's cuda:1 to worker1's cuda:2 >>> >>> rpc.init_rpc( >>> "worker0", >>> rank=0, >>> world_size=2, >>> backend=rpc.BackendType.TENSORPIPE, >>> rpc_backend_options=options >>> ) >>> >>> x = torch.ones(2) >>> rets = rpc.rpc_sync("worker1", add, args=(x.to(0), 1)) >>> # The first argument will be moved to cuda:1 on worker1. When >>> # sending the return value back, it will follow the invert of >>> # the device map, and hence will be moved back to cuda:0 and >>> # cuda:1 on worker0 >>> print(rets[0]) # tensor([2., 2.], device='cuda:0') >>> print(rets[1]) # tensor([2., 2.], device='cuda:1') z=`set_device_map` only supports 1-to-1 mapping, trying to map rrN)rr'r"rr_set_device_map)r)r.rrcurr_device_mapsrrr,rrset_device_mapks - z*TensorPipeRpcBackendOptions.set_device_mapcCst||_dS)ab Set local devices used by the TensorPipe RPC agent. When processing CUDA RPC requests, the TensorPipe RPC agent will properly synchronize CUDA streams for all devices in this ``List``. Args: devices (List of int, str, or torch.device): local devices used by the TensorPipe RPC agent. N)rr)r)rrrr set_devicess z'TensorPipeRpcBackendOptions.set_devices)__name__ __module__ __qualname____doc__ rpc_contantsDEFAULT_NUM_WORKER_THREADSDEFAULT_RPC_TIMEOUT_SECDEFAULT_INIT_METHODintfloatstrrdict DeviceTyperr(r1r2 __classcell__rrr,rr.s4$  :)typingrrr torch._C._distributed_rpcrrr7r;r=rr?__all__rr>rrrrrrrrs