Hnimrama/ix atom multinode

cvs/core/orchestrators/container.py

@@ -618,6 +618,10 @@ def exec(self, cmd, hosts=None, timeout=None, detailed=False):
 
         return self.runtime.exec(self.container_id, cmd, hosts, timeout, detailed)
 
+    def exec_on_host(self, cmd, hosts=None, timeout=None, detailed=False):
+        """Execute command on the cluster host OS (SSH), not inside the container."""
+        return super().exec(cmd, hosts=hosts, timeout=timeout, detailed=detailed)
+
     def exec_on_head(self, cmd, timeout=None):
         """
         Execute command directly on head node (baremetal).

cvs/core/orchestrators/unittests/test_container.py

@@ -15,7 +15,7 @@
 # patched once in setUp (not per method); _make() returns a fresh orch + runtime mock.
 
 import unittest
-from unittest.mock import MagicMock, patch
+from unittest.mock import MagicMock, call, patch
 
 from cvs.core.orchestrators.factory import OrchestratorConfig, _resolve_container_lifetime
 from cvs.core.orchestrators.container import ContainerOrchestrator
@@ -258,6 +258,21 @@ def test_setup_sshd_multinode_attempts_setup(self):
         }
         self.assertTrue(orch.setup_sshd())
         self.assertTrue(runtime.exec.called)
+        cmds = [call.args[0] for call in runtime.exec.call_args_list]
+        self.assertFalse(any("apt-get" in cmd for cmd in cmds))
+
+    @patch("time.sleep", lambda *_a, **_k: None)
+    def test_setup_sshd_multinode_fails_when_sshd_missing(self):
+        orch, runtime = self._make(lifetime="per_run")
+        orch.container_id = "cvs_iter_test"
+
+        def _exec_side_effect(cmd, **kwargs):
+            if "command -v /usr/sbin/sshd" in cmd:
+                return {"10.0.0.1": {"exit_code": 127, "output": "missing sshd"}}
+            return {"10.0.0.1": {"exit_code": 0}}
+
+        runtime.exec.side_effect = _exec_side_effect
+        self.assertFalse(orch.setup_sshd())
 
     # ------------------------------------------------------------------
     # teardown_containers lifetime branching

cvs/input/cluster_file/inferencex_atom_cluster.json

@@ -0,0 +1,35 @@
+{
+  "_comment": "InferenceX ATOM cluster template (container backend). Copy to ~/input/cluster_file/inferencex_atom_cluster.json and edit placeholders. head_node_dict.mgmt_ip MUST be the rank-0 GPU node VPC IP (same as variant params.master_addr for nnodes>1) — not the pytest jumphost. Trim node_dict to one host for single-node variants. Variant config overrides container.image/name/volumes; cluster container block is the fallback default.",
+  "orchestrator": "container",
+  "username": "{user-id}",
+  "priv_key_file": "/home/{user-id}/.ssh/cluster_id_ed25519",
+  "head_node_dict": {
+    "mgmt_ip": "{head-node-ip}"
+  },
+  "env_vars": {},
+  "_env_vars_comment": "Optional host env exported on each GPU node before container setup (e.g. ROCm install on host). Usually empty when the workload image carries ROCm/vLLM.",
+  "node_dict": {
+    "{head-node-ip}": {
+      "bmc_ip": "NA",
+      "vpc_ip": "{head-node-ip}"
+    },
+    "{worker-node-ip}": {
+      "bmc_ip": "NA",
+      "vpc_ip": "{worker-node-ip}"
+    }
+  },
+  "container": {
+    "lifetime": "per_run",
+    "image": "rocm/atom-dev:latest",
+    "name": "inferencex_atom",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G"
+      }
+    }
+  }
+}