fix: avoid HTTP probes on gRPC worker groups

src/dstack/_internal/server/services/jobs/configurators/base.py

@@ -495,6 +495,17 @@ def _probes(self) -> list[ProbeSpec]:
             if probes is not None:
                 return list(map(_probe_config_to_spec, probes))
             # Generate default probe if model is set
+            if self.replica_group_name is not None:
+                replica_group = next(
+                    (
+                        group
+                        for group in self.run_spec.configuration.replica_groups
+                        if group.name == self.replica_group_name
+                    ),
+                    None,
+                )
+                if replica_group is not None and replica_group.router is None:
+                    return []
             model = self.run_spec.configuration.model
             if isinstance(model, OpenAIChatModel):
                 return [_openai_model_probe_spec(model.name, model.prefix)]

src/tests/_internal/server/services/jobs/configurators/test_service.py

@@ -13,6 +13,7 @@
 )
 from dstack._internal.core.models.profiles import SpotPolicy
 from dstack._internal.core.models.resources import Range
+from dstack._internal.core.models.routers import ReplicaGroupRouterConfig
 from dstack._internal.core.models.services import OpenAIChatModel
 from dstack._internal.server.services.docker import ImageConfig
 from dstack._internal.server.services.jobs.configurators.base import get_default_image
@@ -107,6 +108,47 @@ async def test_no_probe_when_no_model(self):
         assert len(job_specs[0].probes) == 0
 
 
+@pytest.mark.windows_only
+def test_default_probe_only_for_router_replica_group():
+    """A model probe should not make gRPC worker groups fail readiness."""
+    configuration = ServiceConfiguration(
+        port=80,
+        image="debian",
+        model=OpenAIChatModel(
+            name="meta-llama/Meta-Llama-3.1-8B-Instruct",
+            format="openai",
+        ),
+        replicas=[
+            ReplicaGroup(
+                name="router",
+                replicas=1,
+                router=ReplicaGroupRouterConfig(),
+            ),
+            ReplicaGroup(name="worker", replicas=1),
+        ],
+    )
+    run_spec = get_run_spec(run_name="run", repo_id="id", configuration=configuration)
+
+    router_probes = ServiceJobConfigurator(run_spec, replica_group_name="router")._probes()
+    worker_probes = ServiceJobConfigurator(run_spec, replica_group_name="worker")._probes()
+
+    assert len(router_probes) == 1
+    assert worker_probes == []
+
+    explicit_configuration = configuration.model_copy(
+        update={"probes": [ProbeConfig(type="http", url="/health")]}
+    )
+    explicit_run_spec = get_run_spec(
+        run_name="run", repo_id="id", configuration=explicit_configuration
+    )
+    explicit_worker_probes = ServiceJobConfigurator(
+        explicit_run_spec, replica_group_name="worker"
+    )._probes()
+
+    assert len(explicit_worker_probes) == 1
+    assert explicit_worker_probes[0].url == "/health"
+
+
 def _make_run_spec(replicas, **service_kwargs):
     configuration = ServiceConfiguration(
         port=80,