feature: add RBLNLMCacheConnectorV1 for LMCache KV cache offloading (#523)

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: rebel-jinhwan

vllm_rbln/distributed/kv_transfer/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

vllm_rbln/distributed/kv_transfer/kv_connector/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

vllm_rbln/distributed/kv_transfer/kv_connector/factory.py

@@ -19,3 +19,8 @@
     "vllm_rbln.distributed.kv_transfer.kv_connector.v1.rbln_nixl_connector",
     "RblnNixlConnector",
 )
+KVConnectorFactory.register_connector(
+    "RBLNLMCacheConnectorV1",
+    "vllm_rbln.distributed.kv_transfer.kv_connector.v1.rbln_lmcache_connector",
+    "RBLNLMCacheConnectorV1",
+)

vllm_rbln/distributed/kv_transfer/kv_connector/v1/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

vllm_rbln/distributed/kv_transfer/kv_connector/v1/rbln_lmcache_connector.py

@@ -0,0 +1,33 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""RBLN LMCache KV connector for vLLM — thin re-export shim.
+
+All RBLN-specific logic lives in the ``lmcache_rbln`` package.
+This module re-exports ``RBLNLMCacheConnectorV1`` so that vllm-rbln's
+factory registration can reference it via a stable path inside the
+``vllm_rbln`` namespace.
+
+Usage in vLLM config::
+
+    kv_connector = (
+        "vllm_rbln.distributed.kv_transfer"
+        ".kv_connector.v1.rbln_lmcache_connector"
+        ".RBLNLMCacheConnectorV1"
+    )
+"""
+
+from lmcache_rbln.integration.vllm.connector import RBLNLMCacheConnectorV1
+
+__all__ = ["RBLNLMCacheConnectorV1"]

vllm_rbln/v1/worker/rbln_model_runner.py

@@ -4510,6 +4510,9 @@ def rbln_copy_kv_blocks(
 
             kv_transfer_group.set_host_xfer_buffer_ops(rbln_copy_kv_blocks)
 
+            if hasattr(kv_transfer_group, "set_runtime_holder"):
+                kv_transfer_group.set_runtime_holder(self.runtime_holder)
+
         if self.dcp_world_size > 1:
             layer_type = cast(type[Any], AttentionLayerBase)
             layers = get_layers_from_vllm_config(self.vllm_config, layer_type)