[KV Connector] Update lmcache kv_offloading_backend to use LMCacheMPConnector (#42865)

Signed-off-by: baoloongmao <baoloongmao@tencent.com>

tests/v1/kv_connector/unit/test_config.py

@@ -6,25 +6,56 @@
 import pytest
 
 from vllm.config import CacheConfig, KVTransferConfig, ParallelConfig, VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.factory import KVConnectorFactory
 
 pytestmark = pytest.mark.cpu_test
 
 
+class _StubLMCacheMPConnector:
+    """Stand-in for LMCacheMPConnector used in config-translation tests.
+
+    The real connector module hard-imports the optional ``lmcache`` package
+    at module load time, which is not installed in the cpu_test image. This
+    test only asserts on the connector *name* and the ``extra_config`` dict
+    produced by ``VllmConfig``, never instantiates the connector, so a bare
+    placeholder class is sufficient. Not subclassing ``SupportsHMA`` mirrors
+    the real connector's HMA support (it does not support HMA either)."""
+
+
+@pytest.fixture
+def stub_lmcache_mp_connector(monkeypatch):
+    """Replace the lazy loader so VllmConfig.__post_init__ does not import
+    ``lmcache_mp_connector`` (and thus ``lmcache``) during config tests."""
+    monkeypatch.setitem(
+        KVConnectorFactory._registry,
+        "LMCacheMPConnector",
+        lambda: _StubLMCacheMPConnector,
+    )
+
+
 @pytest.mark.parametrize(
     "kv_offloading_backend,kv_offloading_size,tp,pp,expected_backend,expected_bytes",
     [
         ("native", 4.0, 1, 1, "OffloadingConnector", 4.0 * (1 << 30)),
         # bytes per rank: 8.0 GiB / (2 * 2) = 2.0 GiB
         ("native", 8.0, 2, 2, "OffloadingConnector", 8.0 * (1 << 30)),
-        ("lmcache", 4.0, 1, 1, "LMCacheConnectorV1", 4.0),
-        # size per rank: 8.0 GiB / (2 * 2) = 2.0 GiB
-        ("lmcache", 8.0, 2, 2, "LMCacheConnectorV1", 2.0),
+        # ``lmcache`` backend now defaults to LMCacheMPConnector. The KV
+        # storage capacity is owned by the standalone LMCache server, so
+        # ``kv_offloading_size`` is intentionally not propagated.
+        ("lmcache", 4.0, 1, 1, "LMCacheMPConnector", None),
+        ("lmcache", 8.0, 2, 2, "LMCacheMPConnector", None),
         # When kv_offloading_size is None, offloading is disabled (backend is ignored)
         ("native", None, 1, 1, None, None),
     ],
 )
 def test_kv_connector(
-    kv_offloading_backend, kv_offloading_size, tp, pp, expected_backend, expected_bytes
+    stub_lmcache_mp_connector,
+    kv_offloading_backend,
+    kv_offloading_size,
+    tp,
+    pp,
+    expected_backend,
+    expected_bytes,
 ):
     kv_transfer_config = (
         KVTransferConfig(kv_connector_extra_config={"existing_key": "existing_value"})
@@ -59,10 +90,12 @@ def test_kv_connector(
         # Existing config should be preserved
         assert kv_connector_extra_config["existing_key"] == "existing_value"
     elif kv_offloading_backend == "lmcache":
-        assert kv_connector_extra_config["lmcache.local_cpu"] is True
-        assert kv_connector_extra_config["lmcache.max_local_cpu_size"] == expected_bytes
-        # Existing config should be replaced
-        assert "existing_key" not in kv_connector_extra_config
+        # MP mode does not push lmcache.local_cpu / max_local_cpu_size into
+        # extra config (the LMCache server owns capacity). Pre-existing
+        # extra config entries are preserved as-is.
+        assert "lmcache.local_cpu" not in kv_connector_extra_config
+        assert "lmcache.max_local_cpu_size" not in kv_connector_extra_config
+        assert kv_connector_extra_config["existing_key"] == "existing_value"
 
 
 def _build_config(

vllm/config/vllm.py

@@ -780,10 +780,6 @@ class VllmConfig:
         # If no KVTransferConfig is provided, create a default one.
         if self.kv_transfer_config is None:
             self.kv_transfer_config = KVTransferConfig()
-        num_kv_ranks = (
-            self.parallel_config.tensor_parallel_size
-            * self.parallel_config.pipeline_parallel_size
-        )
 
         if kv_offloading_backend == "native":
             if envs.VLLM_USE_SIMPLE_KV_OFFLOAD:
@@ -795,12 +791,12 @@ class VllmConfig:
                 {"cpu_bytes_to_use": kv_offloading_size * (1 << 30)}
             )
         elif kv_offloading_backend == "lmcache":
-            self.kv_transfer_config.kv_connector = "LMCacheConnectorV1"
-            kv_gb_per_rank = kv_offloading_size / num_kv_ranks
-            self.kv_transfer_config.kv_connector_extra_config = {
-                "lmcache.local_cpu": True,
-                "lmcache.max_local_cpu_size": kv_gb_per_rank,
-            }
+            # Default to LMCache multi-process (MP) mode. The actual KV
+            # storage capacity is managed by the standalone LMCache server
+            # process, so ``kv_offloading_size`` is not propagated here.
+            # ``LMCacheMPConnector`` falls back to ``tcp://localhost:5555``
+            # when host/port are not provided via extra_config.
+            self.kv_transfer_config.kv_connector = "LMCacheMPConnector"
 
         # This is the same for all backends
         self.kv_transfer_config.kv_role = "kv_both"