Request Transport

LMCache MP clients select the request transport from the LMCache server URL. This transport carries control and cache-operation requests; it is separate from --supported-transfer-mode, which controls how KV data moves between the engine worker and the LMCache server.

Transport schemes

Endpoint	Transport	Status
host:port or tcp://host:port	ZMQ over TCP	Supported. A URL without a scheme defaults to tcp://.
ipc://path or inproc://name	ZMQ IPC or in-process	Recognized by the client factory; the standard LMCache server currently binds ZMQ over TCP.
grpc://host:port or grpc+unix:///path	gRPC	Supported.

gRPC implementation and schema development

The gRPC client and server live under lmcache/v1/multiprocess/transport/grpc_impl. They expose the same named RequestClient operations and annotated engine-module handlers as ZMQ, while protobuf defines the gRPC wire format. Schemas live in the sibling protos directory, and package builds generate their Python bindings under _proto_gen.

After changing a schema, regenerate and validate all bindings with:

pip install -r requirements/proto.txt
python -m lmcache.v1.multiprocess.transport.grpc_impl._proto_gen._generate

For a single vLLM connector, set the scheme in lmcache.mp.host and keep the port in lmcache.mp.port. A ZMQ configuration is:

{
  "lmcache.mp.host": "tcp://localhost",
  "lmcache.mp.port": 5555
}

To select gRPC, use the same configuration shape with a grpc:// host:

{
  "lmcache.mp.host": "grpc://localhost",
  "lmcache.mp.port": 5555
}

Start the server with the matching request transport:

lmcache server --transport zmq --host localhost --port 5555
lmcache server --transport grpc --host localhost --port 5555

For multiple servers, specify the scheme on every entry in lmcache.mp.server_urls, for example tcp://host1:5555,tcp://host2:5555 or grpc://host1:5555,grpc://host2:5555. All clients and servers in a deployment must use matching transports.