feat: Maru storage backend to feat/maru-connector

docs/source/kv_cache/storage_backends/maru.rst

@@ -43,14 +43,13 @@ Deploy Model With Maru
 .. code-block:: yaml
 
     chunk_size: 256
-    local_cpu: True
-    max_local_cpu_size: 5
-    remote_url: "maru://localhost:5555"
-    remote_serde: "naive"
+    local_cpu: False
+    max_local_cpu_size: 0
+    save_unfull_chunk: True
 
-    extra_config:
-      maru_pool_size: "4G"
-      save_chunk_meta: False
+    # Maru backend
+    maru_path: "maru://localhost:5555"
+    maru_pool_size: 4
 
 **3. Start vLLM with Maru**
 
@@ -75,11 +74,14 @@ Configuration
    * - Parameter
      - Default
      - Description
-   * - ``remote_url``
+   * - ``maru_path``
      - Required
      - Maru server URL (format: ``maru://host:port``)
+   * - ``maru_pool_size``
+     - ``4.0``
+     - CXL memory pool size per instance in GB (e.g., ``4``, ``0.5``)
 
-**Maru Parameters (via extra_config):**
+**Advanced Parameters (via extra_config):**
 
 .. list-table::
    :header-rows: 1
@@ -88,17 +90,11 @@ Configuration
    * - Parameter
      - Default
      - Description
-   * - ``maru_pool_size``
-     - ``"1G"``
-     - CXL memory pool size per instance (e.g., ``"4G"``, ``"500M"``)
    * - ``maru_instance_id``
      - auto UUID
      - Unique client instance identifier
-   * - ``maru_operation_timeout``
-     - 10.0
-     - Per-operation timeout in seconds
    * - ``maru_timeout_ms``
-     - 2000
+     - 5000
      - ZMQ RPC socket timeout in milliseconds
    * - ``maru_use_async_rpc``
      - true

lmcache/v1/config.py

@@ -236,6 +236,13 @@
         "default": None,
         "env_converter": int,
     },
+    # Maru CXL shared memory backend
+    "maru_path": {"type": Optional[str], "default": None, "env_converter": str},
+    "maru_pool_size": {
+        "type": float,
+        "default": 4.0,
+        "env_converter": float,
+    },
     # Other configurations
     # (Deprecated) The url of the actual remote lmcache instance for auditing.
     # Please use extra_config['audit_actual_remote_url'] instead.

lmcache/v1/storage_backend/__init__.py

@@ -218,6 +218,20 @@ def CreateStorageBackends(
         )
         storage_backends[str(gds_backend)] = gds_backend
 
+    if config.maru_path is not None and "MaruBackend" not in _skip:
+        try:
+            # First Party
+            from lmcache.v1.storage_backend.maru_backend import MaruBackend
+        except ImportError as e:
+            raise ImportError(
+                "The 'maru' and 'maru_lmcache' packages are required "
+                "to use MaruBackend. Please install them according to "
+                "the Maru setup documentation."
+            ) from e
+
+        maru_backend = MaruBackend(config, metadata, loop, dst_device)
+        storage_backends[str(maru_backend)] = maru_backend
+
     if config.remote_url is not None and "RemoteBackend" not in _skip:
         assert local_cpu_backend is not None, (
             "Remote backend requires local CPU backend as a buffer."