Class ClusterFollower

java.lang.Object

smile.llm.cluster.ClusterFollower

All Implemented Interfaces:

AutoCloseable

public final class ClusterFollower
extends Object
implements AutoCloseable

The follower side of a multi-node model: connects to a ClusterLeader, learns its ranks and the NCCL unique id, then executes every message the leader sends, in order, acknowledging each. See ClusterLeader for the wire format.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static interface	ClusterFollower.Handler	Executes one message; the follower acknowledges it afterwards.
static final record	ClusterFollower.Welcome	What the leader tells a follower on connection.

Method Summary

Modifier and Type	Method	Description
void	close()
static ClusterFollower	connect(String host, int port, int nodeIndex, int ranks, long timeoutMs)	Connects to the leader (retrying until it listens) and completes the handshake.
void	run(ClusterFollower.Handler handler)	Executes the leader's messages until it shuts the cluster down or disconnects.
ClusterFollower.Welcome	welcome()	Returns what the leader assigned.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

connect

public static ClusterFollower connect(String host,
 int port,
 int nodeIndex,
 int ranks,
 long timeoutMs)
                               throws IOException

Connects to the leader (retrying until it listens) and completes the handshake.

Parameters:

host - the leader's host.

port - the leader's cluster port.

nodeIndex - this follower's node number (1, 2, ...; ranks are assigned in node order).

ranks - number of ranks (GPUs) this follower owns.

timeoutMs - how long to keep retrying.

Returns:

the connected follower.

Throws:

IOException - if the leader cannot be reached or the handshake is wrong.

welcome

public ClusterFollower.Welcome welcome()

Returns what the leader assigned.

Returns:

what the leader assigned.

run

public void run(ClusterFollower.Handler handler)
         throws Exception

Executes the leader's messages until it shuts the cluster down or disconnects. A failing message is reported to the leader and the exception is rethrown (the caller should then exit: the other ranks cannot continue).

Parameters:

handler - executes the messages.

Throws:

Exception - the first failure of handler.

close

public void close()

Specified by:

close in interface AutoCloseable