Summary

MultiConnector.__init__ instantiates its child connectors by calling connector_cls(...) directly, bypassing the backward-compat shim that KVConnectorFactory.create_connector / _get_connector_class_with_compat added in #27887. As a result, any legacy connector whose __init__ still has the pre-#27887 signature (self, vllm_config, role) fails with TypeError the moment it is placed inside a MultiConnector, even though the same connector works fine when used as a top-level connector (where it only produces the expected deprecated signature warning).

Environment

• vLLM: 0.18.0 (also reproduced against current main, commit bf45e6d0a)
• Reproducing connector: LMCacheAscendConnectorV1Dynamic from LMCache-Ascend (tracks vLLM ≤ 0.11.0, still uses the 2-arg signature)

Repro

--kv-transfer-config with a MultiConnector that wraps a legacy 2-arg connector:

```json { "kv_connector": "MultiConnector", "kv_role": "kv_producer", "engine_id": "0", "kv_connector_extra_config": { "connectors": [ {"kv_connector": "LMCacheAscendConnectorV1Dynamic", "kv_role": "kv_both", "kv_connector_module_path": "lmcache_ascend.integration.vllm.lmcache_ascend_connector_v1", "kv_connector_extra_config": {"lmcache_rpc_port": "lmcache_p"}}, {"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000"} ] } } ```

Traceback:

``` File ".../vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py", line 166, in init self._connectors.append(connector_cls(temp_config, role, kv_cache_config)) TypeError: LMCacheAscendConnectorV1Dynamic.init() takes 3 positional arguments but 4 were given ```

Right before the TypeError, the factory-level shim does log the expected deprecation warning for the same class:

``` WARNING factory.py:126 Connector LMCacheAscendConnectorV1Dynamic uses deprecated signature with 2 required arguments. Please update to include kv_cache_config as the second argument. ```

So the compat machinery exists and fires correctly at the factory level; MultiConnector just does not use it.

Root cause

#27887 added kv_cache_config as a third __init__ argument and, in factory.py, a two-step shim:

1. _get_connector_class_with_compat probes the class with supports_kw(connector_cls, \"kv_cache_config\"), returns compat_sig=True for old-style classes, emits the deprecation warning.
2. create_connector branches on that flag and calls either connector_cls(config, role) or connector_cls(config, role, kv_cache_config).

The same PR updated multi_connector.py to forward 3 args to children but did not route children through create_connector / the shim. The current child-instantiation line is:

```python

vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py:166

self._connectors.append(connector_cls(temp_config, role, kv_cache_config)) ```

Proposed fix

Route child-connector construction through KVConnectorFactory.create_connector so the shim applies uniformly:

```python from vllm.distributed.kv_transfer.kv_connector.factory import KVConnectorFactory

instead of: connector_cls(temp_config, role, kv_cache_config)

self._connectors.append( KVConnectorFactory.create_connector(temp_config, role, kv_cache_config) ) ```

This has two side effects to consider, both benign in practice:

• an extra Creating v1 connector ... log line per child (informational);
• the hma_enabled check inside create_connector will run, but when --kv-transfer-config is set vLLM already disables the hybrid KV cache manager, so the check does not fire.

Happy to send a PR if the direction is acceptable.

Workaround

Patch the legacy connector's signature locally to accept kv_cache_config=None. This keeps LMCache-Ascend working until upstream adapters catch up with 0.11.1+.