Find each node's management interface, and lift the agent's memlock limit

The agent's defaults said eth0 before normalize looked for the interface
holding the management address, so it never looked: a node on ens3 or
bond0 had NCCL and Gloo bootstrap over an interface it does not have. It
is now found, with eth0 only as the last resort.

The agent unit sets LimitMEMLOCK=infinity. RDMA pins what it registers,
and under systemd's 8 MB default an engine without the capability that
lets root ignore the limit fails its first collective over InfiniBand.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

deploy/bootstrap/onboard-node.sh

@@ -226,6 +226,10 @@ ExecStart=/usr/local/bin/llmgw agent --config /etc/llmgw/agent.yaml
 KillMode=process
 Restart=on-failure
 RestartSec=2s
+# RDMA pins the buffers it registers, and NCCL registers a great deal: with
+# systemd's default 8 MB lock limit, an engine that drops the capability that
+# lets root ignore it fails its first collective over InfiniBand.
+LimitMEMLOCK=infinity
 StateDirectory=llmgw-agent
 
 [Install]

deploy/systemd/llmgw-agent.service

@@ -11,6 +11,10 @@ ExecStart=/usr/local/bin/llmgw agent --config /etc/llmgw/agent.yaml
 KillMode=process
 Restart=on-failure
 RestartSec=2s
+# RDMA pins the buffers it registers, and NCCL registers a great deal: with
+# systemd's default 8 MB lock limit, an engine that drops the capability that
+# lets root ignore it fails its first collective over InfiniBand.
+LimitMEMLOCK=infinity
 StateDirectory=llmgw-agent
 
 [Install]

internal/agent/agent.go

@@ -118,8 +118,11 @@ type Config struct {
 
 // DefaultConfig returns sensible defaults.
 func DefaultConfig() Config {
+	// No MgmtInterface: normalize finds the one holding the management
+	// address, and falls back to eth0 only when it cannot. Defaulted here, it
+	// was never found — every node said eth0 unless told otherwise, and a node
+	// on ens3 or bond0 had NCCL bootstrap over an interface it does not have.
 	return Config{
-		MgmtInterface:     "eth0",
 		HeartbeatInterval: 5 * time.Second,
 		ReconnectBase:     time.Second,
 		ReconnectMax:      30 * time.Second,
@@ -159,7 +162,7 @@ func (c *Config) normalize() error {
 		c.MgmtInterface = interfaceOf(c.MgmtAddr)
 	}
 	if c.MgmtInterface == "" {
-		c.MgmtInterface = d.MgmtInterface
+		c.MgmtInterface = "eth0"
 	}
 	// The directory downloads land in is always one that is looked in. When
 	// they were two settings, a node could fill its disk with fetched weights

internal/agent/agent_test.go

@@ -813,3 +813,18 @@ func TestAShardUsesThisNodesInterfaceAndRoCEIndex(t *testing.T) {
 		t.Error("an InfiniBand shard was given a GID index")
 	}
 }
+
+func TestTheManagementInterfaceIsFoundNotAssumed(t *testing.T) {
+	// It was defaulted to eth0 before it could be looked for, so it never
+	// was: a node on ens3 or bond0 had NCCL bootstrap over nothing.
+	cfg := DefaultConfig()
+	cfg.Server = "http://127.0.0.1:7443"
+	cfg.NodeID = "n1"
+	cfg.MgmtAddr = "127.0.0.1"
+	if err := cfg.normalize(); err != nil {
+		t.Fatal(err)
+	}
+	if cfg.MgmtInterface == "" || cfg.MgmtInterface == "eth0" {
+		t.Errorf("interface = %q, want the loopback interface holding 127.0.0.1", cfg.MgmtInterface)
+	}
+}

internal/cli/serverinit_cluster.go

@@ -62,6 +62,10 @@ ExecStart=/usr/local/bin/llmgw agent --config /etc/llmgw/agent.yaml
 KillMode=process
 Restart=on-failure
 RestartSec=2s
+# RDMA pins the buffers it registers, and NCCL registers a great deal: with
+# systemd's default 8 MB lock limit, an engine that drops the capability that
+# lets root ignore it fails its first collective over InfiniBand.
+LimitMEMLOCK=infinity
 StateDirectory=llmgw-agent
 
 [Install]