def connect_to_remote_host(remote_host=None, job_name="worker"): """Connects to a single machine to enable remote execution on it. Will make devices on the remote host available to use. Note that calling this more than once will work, but will invalidate any tensor handles on the old remote devices. Using the default job_name of worker, you can schedule ops to run remotely as follows: ```python # Enable eager execution, and connect to the remote host. tf.compat.v1.enable_eager_execution() tf.contrib.eager.connect_to_remote_host("exampleaddr.com:9876") with ops.device("job:worker/replica:0/task:1/device:CPU:0"): # The following tensors should be resident on the remote device, and the op # will also execute remotely. x1 = array_ops.ones([2, 2]) x2 = array_ops.ones([2, 2]) y = math_ops.matmul(x1, x2) ``` Args: remote_host: a single or a list the remote server addr in host-port format. job_name: The job name under which the new server will be accessible. Raises: ValueError: if remote_host is None. """ if not remote_host: raise ValueError("Must provide at least one remote_host") remote_host = nest.flatten(remote_host) grpc_prefix = "grpc://" local_port = pywrap_tensorflow.TF_PickUnusedPortOrDie() cluster_def = ClusterDef() job_def = cluster_def.job.add() job_def.name = "localhost" # TODO(fishx): Update this to make sure remote worker has valid ip address # to connect with local. job_def.tasks[0] = "localhost:{}".format(local_port) job_def = cluster_def.job.add() job_def.name = job_name for i in range(len(remote_host)): if remote_host[i].startswith(grpc_prefix): job_def.tasks[i] = remote_host[i][len(grpc_prefix):] else: job_def.tasks[i] = remote_host[i] server_def = ServerDef(cluster=cluster_def, job_name="localhost", task_index=0, protocol="grpc") # TODO(nareshmodi): Make this default since it works in more situations. os.environ["TF_EAGER_REMOTE_USE_SEND_TENSOR_RPC"] = "1" context.set_server_def(server_def)
def setUp(self): # Start the local server. local_port = pywrap_tensorflow.TF_PickUnusedPortOrDie() context.set_server_def(server_def=get_server_def( JOB_NAME, local_server_port=local_port, remote_server_addresses=[ self._cached_server1_target, self._cached_server2_target ], task_index=0))
def connect_to_cluster(cluster_spec_or_resolver, job_name="localhost", task_index=0, protocol=None): """Connects to the given cluster. Will make devices on the cluster available to use. Note that calling this more than once will work, but will invalidate any tensor handles on the old remote devices. If the given local job name is not present in the cluster specification, it will be automatically added, using an unused port on the localhost. Args: cluster_spec_or_resolver: A `ClusterSpec` or `ClusterResolver` describing the cluster. job_name: The name of the local job. task_index: The local task index. protocol: The communication protocol, such as `"grpc"`. If unspecified, will use the default from `python/platform/remote_utils.py`. """ protocol = protocol or remote_utils.get_default_communication_protocol() if isinstance(cluster_spec_or_resolver, server_lib.ClusterSpec): cluster_spec = cluster_spec_or_resolver elif isinstance(cluster_spec_or_resolver, cluster_resolver.ClusterResolver): cluster_spec = cluster_spec_or_resolver.cluster_spec() else: raise ValueError( "`cluster_spec_or_resolver` must be a `ClusterSpec` or a " "`ClusterResolver`.") cluster_def = cluster_spec.as_cluster_def() # Automatically add local job, if not part of the cluster spec. if job_name not in cluster_spec.jobs: local_port = pywrap_tensorflow.TF_PickUnusedPortOrDie() job_def = cluster_def.job.add() job_def.name = job_name # TODO(fishx): Update this to make sure remote worker has valid ip address # to connect with local. job_def.tasks[0] = "localhost:{}".format(local_port) server_def = ServerDef(cluster=cluster_def, job_name=job_name, task_index=task_index, protocol=protocol) # TODO(nareshmodi): Make this default since it works in more situations. os.environ["TF_EAGER_REMOTE_USE_SEND_TENSOR_RPC"] = "1" context.set_server_def(server_def)
def connect_to_cluster(cluster_spec_or_resolver, job_name="localhost", task_index=0, protocol=None, make_master_device_default=True): """Connects to the given cluster. Will make devices on the cluster available to use. Note that calling this more than once will work, but will invalidate any tensor handles on the old remote devices. If the given local job name is not present in the cluster specification, it will be automatically added, using an unused port on the localhost. Args: cluster_spec_or_resolver: A `ClusterSpec` or `ClusterResolver` describing the cluster. job_name: The name of the local job. task_index: The local task index. protocol: The communication protocol, such as `"grpc"`. If unspecified, will use the default from `python/platform/remote_utils.py`. make_master_device_default: If True and a cluster resolver is passed, will automatically enter the master task device scope, which indicates the master becomes the default device to run ops. It won't do anything if a cluster spec is passed. Will throw an error if the caller is currently already in some device scope. """ protocol = protocol or remote_utils.get_default_communication_protocol() if isinstance(cluster_spec_or_resolver, server_lib.ClusterSpec): cluster_spec = cluster_spec_or_resolver elif isinstance(cluster_spec_or_resolver, cluster_resolver.ClusterResolver): cluster_spec = cluster_spec_or_resolver.cluster_spec() else: raise ValueError( "`cluster_spec_or_resolver` must be a `ClusterSpec` or a " "`ClusterResolver`.") cluster_def = cluster_spec.as_cluster_def() # Automatically add local job, if not part of the cluster spec. if job_name not in cluster_spec.jobs: local_port = pywrap_tensorflow.TF_PickUnusedPortOrDie() job_def = cluster_def.job.add() job_def.name = job_name # TODO(fishx): Update this to make sure remote worker has valid ip address # to connect with local. job_def.tasks[0] = "localhost:{}".format(local_port) server_def = ServerDef(cluster=cluster_def, job_name=job_name, task_index=task_index, protocol=protocol) # TODO(nareshmodi): Make this default since it works in more situations. os.environ["TF_EAGER_REMOTE_USE_SEND_TENSOR_RPC"] = "1" context.set_server_def(server_def) if make_master_device_default and isinstance( cluster_spec_or_resolver, cluster_resolver.ClusterResolver ) and cluster_spec_or_resolver.master(): master = cluster_spec_or_resolver.master() master_job_name = None master_task_id = None for job_name in cluster_spec.jobs: for task_id in cluster_spec.task_indices(job_name): task_address = cluster_spec.task_address(job_name, task_id) if master in task_address or task_address in master: master_job_name = job_name master_task_id = task_id break if not master_job_name: raise ValueError( "`make_master_device_default` is set to True but cannot find " "master %s in the cluster" % master) master_device = "/job:{}/replica:0/task:{}".format( master_job_name, master_task_id) if not _device_stack_is_empty(): raise ValueError( "`connect_to_cluster` should not be called inside " "an existing device scope") logging.info("Entering into master device scope: %s", master_device) # TODO(b/138389076): Think of the entering device scope behavior in the # failure recovery case when dealing with preemptions. ops.device(master_device).__enter__()
def connect_to_cluster(cluster_spec_or_resolver, job_name="localhost", task_index=0, protocol=None, make_master_device_default=True): """Connects to the given cluster. Will make devices on the cluster available to use. Note that calling this more than once will work, but will invalidate any tensor handles on the old remote devices. If the given local job name is not present in the cluster specification, it will be automatically added, using an unused port on the localhost. Args: cluster_spec_or_resolver: A `ClusterSpec` or `ClusterResolver` describing the cluster. job_name: The name of the local job. task_index: The local task index. protocol: The communication protocol, such as `"grpc"`. If unspecified, will use the default from `python/platform/remote_utils.py`. make_master_device_default: If True and a cluster resolver is passed, will automatically enter the master task device scope, which indicates the master becomes the default device to run ops. It won't do anything if a cluster spec is passed. Will throw an error if the caller is currently already in some device scope. """ if not context.executing_eagerly(): raise ValueError( "`tf.config.experimental_connect_to_cluster` can only be called in " "eager mode.") protocol = protocol or remote_utils.get_default_communication_protocol() if isinstance(cluster_spec_or_resolver, server_lib.ClusterSpec): cluster_spec = cluster_spec_or_resolver elif isinstance(cluster_spec_or_resolver, cluster_resolver.ClusterResolver): if cluster_spec_or_resolver.master() in _LOCAL_MASTERS: # Do nothing if the master is local. return cluster_spec = cluster_spec_or_resolver.cluster_spec() else: raise ValueError( "`cluster_spec_or_resolver` must be a `ClusterSpec` or a " "`ClusterResolver`.") cluster_def = copy.deepcopy(cluster_spec.as_cluster_def()) # Automatically add local job, if not part of the cluster spec. if job_name not in cluster_spec.jobs: local_port = pywrap_tensorflow.TF_PickUnusedPortOrDie() job_def = cluster_def.job.add() job_def.name = job_name # TODO(fishx): Update this to make sure remote worker has valid ip address # to connect with local. job_def.tasks[0] = "localhost:{}".format(local_port) server_def = ServerDef(cluster=cluster_def, job_name=job_name, task_index=task_index, protocol=protocol, default_session_config=context.context().config) if context.get_server_def() is None: context.set_server_def(server_def) else: context.update_server_def(server_def) if make_master_device_default and isinstance( cluster_spec_or_resolver, cluster_resolver.ClusterResolver ) and cluster_spec_or_resolver.master(): master = cluster_spec_or_resolver.master() master_job_name = None master_task_id = None for job_name in cluster_spec.jobs: for task_id in cluster_spec.task_indices(job_name): task_address = cluster_spec.task_address(job_name, task_id) if master in task_address or task_address in master: master_job_name = job_name master_task_id = task_id break if not master_job_name: raise ValueError( "`make_master_device_default` is set to True but cannot find " "master %s in the cluster" % master) master_device = "/job:{}/replica:0/task:{}".format( master_job_name, master_task_id) master_device = device_util.canonicalize(master_device) current_device = device_util.current() if current_device: current_device = device_util.canonicalize(current_device) if current_device and current_device != master_device: raise ValueError( "`connect_to_cluster` is called inside existing device " "scope %s, which is different from the master device " "scope %s to enter. This is not allowed." % (current_device, master_device)) # TODO(b/138389076): Think of the entering device scope behavior in the # failure recovery case when dealing with preemptions. if not current_device: logging.info("Entering into master device scope: %s", master_device) ops.device(master_device).__enter__()