nvflare.private.fed.server.fed_server module

class BaseServer(project_name=None, min_num_clients=2, max_num_clients=10, heart_beat_timeout=600, handlers: List[FLComponent] | None = None, shutdown_period=30.0)

Bases: ABC

Base server that provides the clients management and server deployment.

client_cleanup()

close()

Shutdown the server.

deploy(args, grpc_args=None, secure_train=False)

Start a grpc server and listening the designated port.

fl_shutdown()

get_all_clients() → Dict[str, Client]

Get the list of registered clients.

Returns:

client}

Return type:

A dict of {client_token

get_cell()

logout_client(token)

notify_dead_client(client)

Called to do further processing of the dead client

Parameters:

client – the dead client

Returns:

abstract remove_client_data(token)

remove_dead_clients()

set_admin_server(admin_server)

class FederatedServer(project_name=None, min_num_clients=2, max_num_clients=10, cmd_modules=None, heart_beat_timeout=600, handlers: List[FLComponent] | None = None, args=None, secure_train=False, snapshot_persistor=None, overseer_agent=None, shutdown_period=30.0, check_engine_frequency=3.0)

Bases: BaseServer

Federated server services.

Parameters:

• project_name – server project name.

• min_num_clients – minimum number of contributors at each round.

• max_num_clients – maximum number of contributors at each round.

• cmd_modules – command modules.

• heart_beat_timeout – heartbeat timeout

• handlers – A list of handler

• args – arguments

• secure_train – whether to use secure communication

abort_run()

authentication_check(request: Message, state_check)

client_heartbeat(request: Message) → Message

close()

Shutdown the server.

create_job_cell(job_id, root_url, parent_url, secure_train, server_config) → Cell

create_run_manager(workspace, job_id)

deploy(args, grpc_args=None, secure_train=False)

Start a grpc server and listening the designated port.

fl_shutdown()

notify_dead_client(client)

Called to do further processing of the dead client

Parameters:

client – the dead client

Returns:

overseer_callback(overseer_agent)

process_job_failure(request: Message)

quit_client(request: Message) → Message

Existing client quits the federated training process.

Server will stop sharing the global model with the client, further contribution will be rejected.

This function does not change min_num_clients and max_num_clients.

register_client(request: Message) → Message

Register new clients on the fly.

Each client must get registered before getting the global model. The server will expect updates from the registered clients for multiple federated rounds.

This function does not change min_num_clients and max_num_clients.

remove_client_data(token)

reset_tokens()

Reset the token set.

After resetting, each client can take a token and start fetching the current global model. This function is not thread-safe.

run_engine()

start_run(job_id, run_root, conf, args, snapshot)

stop_run_engine_cell()

stop_training()

task_meta_info(client_name)

Task meta information.

The model_meta_info uniquely defines the current model, it is used to reject outdated client’s update.