[Bugfix][KVConnector] Fix MultiConnector bypassing compat shim for legacy child connectors

tests/v1/kv_connector/unit/test_backwards_compatibility.py

@@ -273,3 +273,95 @@ def test_signature_detection_with_mocking():
         assert isinstance(new_connector, NewStyleTestConnector)
         assert new_connector._kv_cache_config is not None
         assert new_connector._kv_cache_config == kv_cache_config
+
+
+# ---------------------------------------------------------------------------
+# Regression tests for issue #40690:
+# MultiConnector must route child construction through the compat-aware factory
+# so that legacy 2-arg connectors work as children, not just at the top level.
+# ---------------------------------------------------------------------------
+
+_THIS_MODULE = "tests.v1.kv_connector.unit.test_backwards_compatibility"
+
+
+@pytest.mark.parametrize("role", [KVConnectorRole.SCHEDULER, KVConnectorRole.WORKER])
+def test_multiconnector_old_style_child_instantiation(role):
+    """
+    Regression test for #40690.
+
+    A legacy connector with the old 2-arg __init__(self, vllm_config, role)
+    must work as a child of MultiConnector.  Before the fix, MultiConnector
+    directly called connector_cls(config, role, kv_cache_config), which passed
+    a third positional argument that the old-style connector doesn't accept,
+    raising TypeError.  After the fix, child construction goes through
+    KVConnectorFactory.create_connector, which applies the same compat shim
+    that the top-level path uses.
+    """
+    from vllm.distributed.kv_transfer.kv_connector.v1.multi_connector import (
+        MultiConnector,
+    )
+
+    vllm_config = create_vllm_config(
+        kv_connector="MultiConnector",
+        kv_role="kv_both",
+        kv_connector_extra_config={
+            "connectors": [
+                {
+                    "kv_connector": "OldStyleTestConnector",
+                    "kv_role": "kv_both",
+                    "kv_connector_module_path": _THIS_MODULE,
+                }
+            ]
+        },
+    )
+
+    scheduler = create_scheduler(vllm_config)
+    kv_cache_config = scheduler.kv_cache_config
+
+    # This must not raise TypeError before or after the fix.
+    connector = KVConnectorFactory.create_connector(vllm_config, role, kv_cache_config)
+
+    assert isinstance(connector, MultiConnector)
+    assert len(connector._connectors) == 1
+    child = connector._connectors[0]
+    assert isinstance(child, OldStyleTestConnector)
+    # Old-style connector receives no kv_cache_config
+    assert child._kv_cache_config is None
+
+
+@pytest.mark.parametrize("role", [KVConnectorRole.SCHEDULER, KVConnectorRole.WORKER])
+def test_multiconnector_new_style_child_unaffected(role):
+    """
+    Ensure that the fix does not break new-style 3-arg children inside
+    MultiConnector.
+    """
+    from vllm.distributed.kv_transfer.kv_connector.v1.multi_connector import (
+        MultiConnector,
+    )
+
+    vllm_config = create_vllm_config(
+        kv_connector="MultiConnector",
+        kv_role="kv_both",
+        kv_connector_extra_config={
+            "connectors": [
+                {
+                    "kv_connector": "NewStyleTestConnector",
+                    "kv_role": "kv_both",
+                    "kv_connector_module_path": _THIS_MODULE,
+                }
+            ]
+        },
+    )
+
+    scheduler = create_scheduler(vllm_config)
+    kv_cache_config = scheduler.kv_cache_config
+
+    connector = KVConnectorFactory.create_connector(vllm_config, role, kv_cache_config)
+
+    assert isinstance(connector, MultiConnector)
+    assert len(connector._connectors) == 1
+    child = connector._connectors[0]
+    assert isinstance(child, NewStyleTestConnector)
+    # New-style connector receives kv_cache_config
+    assert child._kv_cache_config is not None
+    assert child._kv_cache_config == kv_cache_config

vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py

@@ -160,10 +160,12 @@ def __init__(
 
         self._connectors: list[KVConnectorBase_V1] = []
         self._ktc_kv_transfer_config = []
-        for connector_cls, temp_config in self._get_connector_classes_and_configs(
-            vllm_config
-        ):
-            self._connectors.append(connector_cls(temp_config, role, kv_cache_config))
+        for temp_config in self._get_child_configs(vllm_config):
+            connector = KVConnectorFactory.create_connector(
+                temp_config, role, kv_cache_config
+            )
+            assert isinstance(connector, KVConnectorBase_V1)
+            self._connectors.append(connector)
             self._ktc_kv_transfer_config.append(temp_config.kv_transfer_config)
 
         # A mapping from request id to the index of the connector chosen to
@@ -183,31 +185,41 @@ def prefer_cross_layer_blocks(self) -> bool:
         return all(c.prefer_cross_layer_blocks for c in self._connectors)
 
     @classmethod
-    def _get_connector_classes_and_configs(
-        cls, vllm_config: "VllmConfig"
-    ) -> list[tuple[type[KVConnectorBaseType], "VllmConfig"]]:
+    def _get_child_configs(cls, vllm_config: "VllmConfig") -> list["VllmConfig"]:
+        """Return one VllmConfig per child connector, with kv_transfer_config
+        set to the child's KVTransferConfig.  Class resolution is intentionally
+        omitted here so that callers such as __init__ can delegate it to
+        KVConnectorFactory.create_connector and avoid resolving the class twice.
+        """
         assert vllm_config.kv_transfer_config is not None
         ktcs = vllm_config.kv_transfer_config.kv_connector_extra_config.get(
             "connectors"
         )
         assert ktcs is not None
-        ret: list[tuple[type[KVConnectorBaseType], VllmConfig]] = []
+        ret: list[VllmConfig] = []
         for ktc in ktcs:
             temp_config = copy.copy(vllm_config)
             engine_id = ktc.get("engine_id", vllm_config.kv_transfer_config.engine_id)
             temp_config.kv_transfer_config = KVTransferConfig(
                 **ktc, engine_id=engine_id
             )
-            ret.append(
-                (
-                    KVConnectorFactory.get_connector_class(
-                        temp_config.kv_transfer_config
-                    ),
-                    temp_config,
-                )
-            )
+            ret.append(temp_config)
         return ret
 
+    @classmethod
+    def _get_connector_classes_and_configs(
+        cls, vllm_config: "VllmConfig"
+    ) -> list[tuple[type[KVConnectorBaseType], "VllmConfig"]]:
+        return [
+            (
+                KVConnectorFactory.get_connector_class(
+                    temp_config.kv_transfer_config  # type: ignore[arg-type]
+                ),
+                temp_config,
+            )
+            for temp_config in cls._get_child_configs(vllm_config)
+        ]
+
     def register_cross_layers_kv_cache(
         self, kv_cache: torch.Tensor, attn_backend: type[AttentionBackend]
     ):