קgddlmZmZmZmZddlZddlmZddlm Z ee e ej fZdgZdedej fd Zd eeefdeej ej ffd Zd eedeej fd ZGddeZdS))DictListOptionalUnionN) _TensorPipeRpcBackendOptionsBase) constantsTensorPipeRpcBackendOptionsdevicereturncttj|}|jdkrtd|jd|S)NcudazA`set_devices` expect a list of CUDA devices, but got device type .)torchr type ValueError)r s Y/var/www/html/ai-engine/env/lib/python3.11/site-packages/torch/distributed/rpc/options.py _to_devicersP \& ! !F {f *!; * * *    M device_mapc i}i}|D]Y\}}tj|tj|}}||vrtd|d||d||||<|||<Z|S)Nz9`device_map` only supports 1-to-1 mapping, trying to map  and  to )itemsrr r)rfull_device_map reverse_mapkvs r_to_device_maprs9;O46K  ""1|A Q1  A!"AA)4QAA=>AA  A rdevicescFttt|S)N)listmapr)r s r_to_device_listr$*s J(( ) ))rceZdZdZejejejddddddede de de e e e e e ffde ee d e ed e effd Zd e d e e e fffd Zdee fdZxZS)r a' The backend options for :class:`~torch.distributed.rpc.TensorPipeAgent`, derived from :class:`~torch.distributed.rpc.RpcBackendOptions`. Args: num_worker_threads (int, optional): The number of threads in the thread-pool used by :class:`~torch.distributed.rpc.TensorPipeAgent` to execute requests (default: 16). rpc_timeout (float, optional): The default timeout, in seconds, for RPC requests (default: 60 seconds). If the RPC has not completed in this timeframe, an exception indicating so will be raised. Callers can override this timeout for individual RPCs in :meth:`~torch.distributed.rpc.rpc_sync` and :meth:`~torch.distributed.rpc.rpc_async` if necessary. init_method (str, optional): The URL to initialize the distributed store used for rendezvous. It takes any value accepted for the same argument of :meth:`~torch.distributed.init_process_group` (default: ``env://``). device_maps (Dict[str, Dict], optional): Device placement mappings from this worker to the callee. Key is the callee worker name and value the dictionary (``Dict`` of ``int``, ``str``, or ``torch.device``) that maps this worker's devices to the callee worker's devices. (default: ``None``) devices (List[int, str, or ``torch.device``], optional): all local CUDA devices used by RPC agent. By Default, it will be initialized to all local devices from its own ``device_maps`` and corresponding devices from its peers' ``device_maps``. When processing CUDA RPC requests, the agent will properly synchronize CUDA streams for all devices in this ``List``. N)num_worker_threads rpc_timeout init_method device_mapsr _transports _channelsr&r'r(r)r r*r+c |ind|D}|gnt|} t||||||| dS)Nc4i|]\}}|t|S)r).0rrs r z8TensorPipeRpcBackendOptions.__init__..^s&GGG41a!^A&&GGGr)rr$super__init__) selfr&r'r(r)r r*r+full_device_mapsfull_device_list __class__s rr2z$TensorPipeRpcBackendOptions.__init__Ps" BGG;3D3D3F3FGGG  ")22og6N6N             rtorc Lt|}tj}||vrZ|D]E\}}|||vr6||||kr$t d|d|d|||Ft||dS)a4 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)rr1r)rr_set_device_map)r3r7rrcurr_device_mapsrrr6s rset_device_mapz*TensorPipeRpcBackendOptions.set_device_mapksZ)44 77. ! ! !'--//  1(,,,6Fr6J16M1M1M$L#$LL*+LL2B22Fq2ILL O44444rc.t||_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)r$r )r3r s r set_devicesz'TensorPipeRpcBackendOptions.set_devicess'w// r)__name__ __module__ __qualname____doc__ rpc_contantsDEFAULT_NUM_WORKER_THREADSDEFAULT_RPC_TIMEOUT_SECDEFAULT_INIT_METHODintfloatstrrr DeviceTyperr2r;r= __classcell__)r6s@rr r .s:H#/"I)A';IM.2&*$(         d3Z-C(D#DEF  $z*+ d^ D>      68585$z:7M2N858585858585t 04 #3 0 0 0 0 0 0 0 0r)typingrrrrrtorch._C._distributed_rpcrr rBrFrHr rI__all__rrr$r r.rrrOs@............ FFFFFF''''''3U\) * ( )zelZ+, %, $%"*T*-*$u|2D****A0A0A0A0A0"BA0A0A0A0A0r